Research Mode (Cafes)

Overview

Research Mode ("Cafes") is a standalone research feature where you create independent research workspaces, gather material from multiple sources, and curate findings for use in your Scribes. Available to every signed-in user — token balance is the only limit.

Getting Started

1. Click Research in the navigation bar
2. Click + New Cafe to create a research workspace
3. Name your Cafe (e.g., "Climate Research", "AI Startups")
4. Start searching and gathering research

Organizing Cafés into Groups (W1.43)

The /research page — where your Cafe cards live — lets you organize cafés into groups ("folders"). A group is pure organization: it doesn't change what's in a café, doesn't cost tokens, and doesn't affect generation. A café can belong to several groups at once — an "Avocados" café can sit in Health, Super foods, and Healthy fats all at the same time.

Creating a group

On desktop, drag one café card onto another to create a group holding both — a dialog asks you to name it (pre-filled "New group"), and the group persists once you click Create. On any device, click a café card's ⋮ menu and choose "Add to group…" — pick an existing group or create a new one from the same picker. This is the only path on a phone, since long-press drag is deliberately not used there (it fights page scroll).

Adding a café to an existing group

Drag a café card onto a folder tile to add it — a small undo toast confirms the add. Or use the same ⋮ → "Add to group…" picker and pick the group. Inside an open group, "+ Add café" lists every café not already a member so you can add more without leaving the group.

Dragging is always a copy, never a move. Adding a café to a group never removes it from anywhere else it already belongs — a café can build up membership in as many groups as you like.

Viewing and managing a group

Groups render as folder tiles beside your café cards on /research — folders always sort first, shown with a folder icon, a count pill, and a small preview of the cafés inside. Click a folder tile to open it. A café that belongs to a group is not duplicated in the top-level grid — find it inside its group, or by search (see below).

Inside an open group:

• Click the group's title to rename it inline.
• "+ Add café" adds more cafés to the group.
• Each member card's ✕ removes that café from this group only — the café itself is untouched and stays in every other group it belongs to.
• The group's own ⋮ → Delete group removes the group. Deleting a group never deletes any of its cafés — the confirmation dialog says so, and names how many cafés will survive. An empty group (one you've removed every café from) simply stays as an empty folder — it isn't auto-deleted.

Finding a café inside a group with search

Search on /research flattens grouped results: if your search term matches a café that lives inside a group, that café shows up as its own card (not buried inside a closed folder), carrying small badges naming every group it's in. A folder tile itself only appears in results when the search term matches the group's name.

Groups are private

Groups are yours alone. They never appear on a shared café — if you share a café via a read-only link, nothing about which group(s) it belongs to (or whether it belongs to any) is visible to the person you shared it with, and there is no way for them to see your groups list at all.

Groups in the cross-café pickers

Wherever you can pull cafés into something else — importing sources from another café or importing tray items into a book chapter from the Workspace — the café picker is now organized by group. A café in two groups is listed under both, so you can find it however you're thinking about your research that day.

No nesting, and it's free

Groups are one level deep — you can't put a group inside a group (dragging a folder onto another folder does nothing). Creating, renaming, adding to, removing from, and deleting a group all cost zero tokens — organizing your cafés is always free.

The Workspace: Three Zones (W1.13)

As of W1.42, opening a Cafe shows three columns side by side — this section describes the shipped, current shell. (It was three "zones" from W1.13; the shape and the labels below are the W1.42 rebuild, so a screenshot or description from before 2026-08-21 will look different — the underlying capabilities described throughout this page are unchanged, only where you find them moved.)

• Left panel — three tabs: Chats (every saved conversation in this Cafe), Reports (everything the AI has written for you — Auto-Research, Auto-Write, Detect Conflicts, Find the Truth, Verify, People research), and Sources — organized as an index of runs (Auto-Research, Auto-Write, Detect Conflicts, Find the Truth, People research, Verify reports, then Individual sources last — see Sources Rail below), redesigned again in W1.14 and joined by People research in W55. As of W1.44, each tab carries a live count next to its label (how many chats, how many distinct reports, how many sources) so you can tell at a glance whether a tab actually has anything in it without opening it first; a tab shows no number while its count is still loading rather than flashing a wrong zero. Click a run to lazy-load its sources as dense rows (not cards); each row shows a type icon, truncated title, status dots (credibility, verify, DOI), and a scope checkbox (see Ask-Your-Cafe Scope below). A footer at the bottom of the panel shows "N of M sources in scope". The panel is drag-resizable at its edge, and collapsible.
• Reading pane (centre) — the single place anything gets read or set up: a source, a Serving Tray item, a research report, or — as of W1.44 — a run's own setup and progress (see below). When nothing is open, this is the Ask Your Cafe chat home — see below. Click any row in the left panel and it opens here, with one action toolbar on top (no more per-row button rows).
• Serving Tray (right) — two tabs once expanded: Tray (the snippets and notes you've deliberately saved) and Cited (new in W1.42 — an automatic, numbered list of everything cited in whatever's open in the centre column right now). Defaults to a slim collapsible strip so it doesn't eat width. Expand it to drag-reorder items; click any item to read/edit it in the pane.

At the top of the workspace, a header carries the Cafe's name, its source count, a conflicts count when there are any, and your token balance. Add sources, Share, and a ⋯ menu (Café settings) sit beside it.

As of W1.44, every research verb launches from a single ✦ Deep button on the chat composer, not from a menu in this header. The Deep research ▾ menu that used to live up here (itself a W1.42 rebuild of six separate buttons) is retired — see Starting a Run: The Deep Launcher below for where it went and what changed about how a run opens.

Reading is URL-addressable. Whatever is open in the pane is reflected in the address bar (?read=...) — refresh the page, use browser Back/Forward, or share the exact URL and the same source/report/tray item opens back up. As of W1.44, a run's setup and its live progress are addressable the same way (?read=setup:... / ?read=run:...), so a run survives a refresh and reattaches exactly where it was. A "← Ask your Cafe" back button in the pane returns you to the chat home view.

On mobile (narrow screens), the left panel is full-width and tapping a row drills into a full-screen reader with a Back button; the tray remains a slide-in drawer opened from a "Tray (N)" button. As of W1.44, starting a run from the café's library home is two taps instead of one — Ask to open the composer, then ✦ — because the phone no longer has a standalone launch button; the same ✦ picker opens once you're in the chat.

Starting a Run: The Deep Launcher (W1.44)

As of W1.44, every research verb — Auto-Research, Auto-Write, Detect Conflicts, Claims, Find the Truth, and Research a Person — launches from one ✦ button sitting last on the chat composer row, right next to Send (order: mode chip · question box · attach · model · settings · ✦ Deep · Send). This replaced the Deep research ▾ menu that used to sit in the workspace header — if you see an older screenshot or an FAQ answer pointing you at that menu, it's describing the pre-W1.44 shell; the menu no longer exists.

• Click ✦ Deep to open a picker with the same three headings as before — Gather (Auto-Research, Research a Person), Interrogate (Detect Conflicts, Claims, Find the Truth), Write (Auto-Write) — and Auto-Research's three depths (Quick research · Solid research · Leave no stone unturned, each subtitled with a rough time, not a source count) inline under its row, one tap away.
• Type your question first, then pick a verb, and it comes with you. Whatever's in the message box becomes Auto-Research's starting topic, or Find the Truth's claim — the two verbs that have a field for it. The other four (Auto-Write, Detect Conflicts, Claims, Research a Person) don't take a typed question; it stays in the box untouched.
• Picking a verb opens its setup filling the centre column — no overlay, no dimming, your chat and reports stay right where they were. Nothing runs and nothing is charged until you press the setup's own Run/Start button.
• A live run shows a pulse on ✦ and its own row in the picker reads "● running — view" — click it to jump straight back to the run exactly where it was, from anywhere in the café (a new chat, a source you opened, wherever you wandered off to).
• Auto-Research, Auto-Write, Detect Conflicts, and Draft Scribe keep a half-filled setup for the rest of your session — leave for the chat or a source and come back, and your draft is still there, with an explicit "Clear form" control if you want to start over. Find the Truth and Research a Person are the two exceptions: both are multi-step wizards, and each opens fresh every time you return to it rather than restoring a half-finished step — leaving and coming back is how you clear them. Nothing already completed is lost either way: a finished Find the Truth check stays reachable under "Past checks."
• A shared-link viewer never sees ✦, and never sees any run surface — if a link to a run's setup or progress view reaches someone who only has view access, it quietly falls back to the café's chat home instead of showing them the owner's controls.

Ask-AI Popover: Highlighting Anywhere in a Cafe

Wherever you're reading in a Cafe — a source card, a Q&A reply, the Serving Tray, an Auto-Write report, or the full-screen report reader — you can act on any piece of text without hunting for a per-surface button.

Opening the popover

• Highlight text (drag-select 3 or more characters) anywhere in the Cafe → a small popover appears near your selection.
• Right-click a source card → the popover opens scoped to the whole source (not just a snippet). Add to Tray and Save to Notebook carry that source's citation reference along, so attribution is never lost.

The four actions

A fifth action: Dictionary

Highlight a single word instead of a phrase and a Dictionary action appears — click it for an instant AI-generated definition, the same dictionary popup used in book Reading Mode. You can save the definition straight to your notebook from the popup.

Where it works

Anything open in the reading pane (a source, a report, or a tray item), Q&A chat replies, and the full-screen report reader (/cafe/[cafeId]/artefact/[artefactId]) all support the popover — it's the same selection handler regardless of which zone you're highlighting in. On the full-screen report reader, Ask AI answers in place — you get an inline answer grounded in that report's sources without leaving the page. Save to notebook and Add to tray also work there; Verify is not available on the standalone reader. On a shared report link, the popover shows only the read-only subset (no Verify, no owner-only actions).

It knows where the highlight came from

Highlight text inside a specific source or report and Ask AI automatically threads that source/report's title into your question — so the answer accounts for which piece of research the passage came from, even if you don't mention it yourself.

Cafe Settings (Citation Style, Filters, Reference Count, and More)

Each Cafe has a Settings panel. As of W1.42, open it from the ⋯ menu in the Cafe's view bar (previously a gear icon in the retired workspace toolbar). Settings are saved per Cafe and take effect on all subsequent searches in that workspace.

Citation Style

Choose from five citation formats: APA 7 (default), APA 6, MLA 9, Chicago Notes-Bibliography, and Harvard Cite Them Right 2025. Enriched source cards in the Cafe show a formatted citation in the style you pick. The style affects how references appear in the card — authors, title, journal, volume, issue, pages, DOI.

Source Tier Filter

Filter results to show only sources from specific tiers: Academic Journal/Paper, Government Source, News Article, Web Article, Social Post, or Wiki Summary. Leave the filter empty (default) to see all tiers. Enable Academic Sources Only as a quick toggle to restrict results to academic-tier sources.

Reference Count

Set how many sources each search returns. Options: Auto (default, resolves to 7), or a fixed count from 1 to 25 — the same ceiling for every signed-in user, no account tier involved.

Enriched Sources Only

When on, search routes hide sources where metadata enrichment returned only a bare URL (no title, author, or date). Useful when you need citable academic sources rather than raw web results.

Web Search

Toggle web search on or off for this Cafe. When off, the Web search mode disappears from the search bar — only Academic search and identifier paste remain available. Default is on.

---

Identifier Paste (every signed-in user)

Instead of searching, you can paste a known identifier. Click the small "Add by identifier" chip in the search bar's info row (next to the model picker) to open a drawer containing the identifier-paste input — along with Cafe Settings and Providers. Paste the identifier into that input and CoffeeScribe auto-detects the type as you type, showing a pill:

The Enrich button is disabled if the type is unrecognised. On submission, the enrichment pipeline resolves the identifier against the appropriate database (Crossref, OpenAlex, PubMed, Open Library, Google Books, or the Crossref Journals API for ISSNs) and adds an enriched source card to your Cafe.

After the metadata enrichment returns, CoffeeScribe automatically attempts to scrape the full article body from the source URL in the background. This is best-effort:

• Open-access articles — full readable body is usually retrieved and stored in the source card.
• Paywalled publishers (Nature, Science, Elsevier, etc.) — the scrape is blocked by a challenge page; the abstract from the enrichment metadata is kept as the content.
• Either way, the 200 response arrives immediately — the body scrape happens in the background and doesn't slow down the identifier flow.

Identifier paste is available to every signed-in user. Cost: a small token charge applies when OpenAlex is reached during enrichment (other resolvers such as Crossref, PubMed, arXiv, Open Library, and Google Books are platform-absorbed at no per-use cost to you). The background body scrape uses the same URL-scrape path as the URL Scraping feature and is charged at the same rate.

---

Source Cards — Enrichment Surface

Every source card in the Cafe now shows enrichment metadata when available:

Metadata Health Badge

A small pill in the card header indicates how complete the enrichment is:

• Full (green) — author, year, journal/publisher, volume, issue, pages, and DOI all present
• Partial (amber) — some fields present; others missing
• URL only (grey) — no structured metadata; bare URL or failed enrichment

Source Tier Label

A small pill shows the source's classified tier: Academic Journal/Paper, Government Source, News Article, Web Article, Social Post, or Wiki Summary. This reflects the source's type, not its quality.

Academic Provider Label

Academic source cards from Auto-Research show an additional provider label — arXiv, PubMed, OpenAlex, or Scholar — so you can tell at a glance which database the source came from, rather than seeing only the generic "Academic" type.

Freshness Badge

A year pill colour-coded by age:

• Green — published within the last 2 years
• Amber — published 3-5 years ago
• Grey — published more than 5 years ago

Hover the badge to see the full publication month and year.

Source-detail Badge

A small badge showing Well-documented, Some research, Limited research, or No details — a quick read of how much we know about a source, shown on every source card. It is not a judgement of whether the source is true, and the weak end is grey, never red — "Limited research" means we've only captured a little about it (often just a link), not that it's fake. It's a cheap, automatic, structural check — no claim verification happens. The read looks at:

• Academic identifier — a DOI, PMID, or arXiv id (even just in the URL, e.g. a PubMed link) marks a peer-reviewed work as authoritative, even before it's fully enriched
• Source tier — academic journal / government source scores higher than a social post
• Author / date presence — a named author and a publication date are positive signals
• Domain reputation — known reputable domains (.gov, .edu, established publishers, doi.org, arxiv.org, PubMed) score higher
• Metadata completeness — richer enrichment reads as more documented

Important: this badge is not the same as "Verified." "Well-documented" means we have a lot of detail about the source — it does not mean a claim in it has been fact-checked. Whether a claim checks out is the separate Verify flow: every source card has a working "Verify claim" button that runs a real cross-reference loop against other sources and returns a verdict (see Verify Claim below). Hover the badge for the reasons behind the rating.

Use the source-detail badge as a quick first-pass signal; always read the source (and Verify important claims) before citing it.

---

Sources Rail — Finding, Organising, Trusting, and Scoping Your Sources

The Sources rail (left side of the Cafe, redesigned again in W1.14 — organized by run — on top of the W1.13/W1.4 foundation) lazy-loads sources by run instead of rendering everything at once, so it stays scannable and responsive even at hundreds or thousands of sources. Click a run to reveal its rows, then click a row to read that source's full detail — enrichment, citation, evidence, everything the old expanded card showed — in the central reading pane. Key capabilities:

Chats — Your Saved Conversations (W1.35)

A Chats section sits at the very top of the Sources rail, above every run — the first thing you see there, because a Cafe's conversations are saved things now, not scratch that disappears when you move on. It lists every chat you've had in this Cafe, newest first, with the one currently open in the reading pane highlighted.

• Every chat you start is saved. Ask a question, and that conversation is a Chats-rail entry from that moment on — nothing is silently thrown away when you start a new one.
• Click a chat to reopen it. The panel switches to that conversation with every turn intact, and whatever chat you were in is automatically archived (moved into Chats, not deleted) so it's never lost.
• Rename inline. Click a chat's rename control, edit the title, then press Enter or click anywhere else to save — either one commits the new title. Press Escape (or click Cancel) to back out without saving.
• Delete a chat from its own row, behind a confirmation, so it can't happen by accident. Deleting the chat you're currently in resets the panel to a clean new chat; deleting any other chat leaves you exactly where you were.
• A brand-new Cafe with nothing but a conversation still shows Chats. The Sources rail is always present for the Cafe's owner, even before you've added a single source, so your very first chat has somewhere to live.
• Owner-only. Shared-link viewers never see a Chats section — see Q&A on Shared Cafes below.

See Conversation Context below for how a chat's title is set, how the AI's memory of a conversation works, and the Use in answers toggle that lets a saved chat feed your Grounded answers.

Organized by Run (W1.14)

The rail's top level is an index of the runs that produced your sources, always in this order: Auto-Research, Auto-Write, Detect Conflicts, Find the Truth, People research (W55 — see Research a Person), Verify reports, then Individual sources last. Each run shows a human label (its topic/question, the person's name, or the report's title) and a source count — computed without loading any source bodies, so the rail stays light even before you open anything.

Click a run to lazy-load its sources: they load underneath the run row a moment later (a brief spinner shows while they fetch), and collapsing hides them again. Re-expanding the same run is instant — its sources are cached, not re-fetched. This is what keeps a Cafe with hundreds or thousands of sources responsive: you never render a flat wall of everything, only the one run you opened.

A run that produced a report or verdict — Auto-Write, Detect Conflicts, Find the Truth, People research, or Verify — shows an "open output" button (file icon) that jumps straight to that report in the reading pane, without needing to expand the run first. Detect Conflicts runs additionally show a link-out to the dedicated conflicts panel and a shortcut to add the conflict report to your Serving Tray.

Auto-Research's extraction cards ("Comprehensive Summary", "Key Findings", "Most Recent", etc.) are full reports too (W1.17). Expand an Auto-Research run to see its reports:[] list — each card opens in the reading pane with the same full action set as every other report type: Open full, Share, Add to Tray, Delete, Copy, Download .md, Download PDF, and the citation toggle. This closes what used to be the one exception in the whole rail: these cards previously opened with only the bare tray-item chrome (Copy / Edit) and could not be added to the tray, shared, or downloaded as PDF. Historical extraction cards work the same way — the full back-catalogue was converted, not just new runs.

Individual sources holds anything not produced by a run — manually added sources, one-off searches, and older sources from before this feature existed (see Earlier Research below). It's always the last section, and — unlike the other sections — it isn't itself a deletable run.

Individual Sources, Grouped by Kind (W1.25)

As of W1.25, Individual sources is no longer one undifferentiated bucket. It's split into five sub-groups, in this order: YouTube · Links · Uploads · Papers · Other sources — so a Cafe with a hundred manually-added sources reads as five scannable piles instead of one long scroll. Each sub-group works exactly like a run: click it to lazy-load its rows, click again to collapse; expanding it a second time is instant (cached).

Extraction cards nest under their source, one open at a time. A source that has extraction cards (Summary, Key Points, Fact Check, etc.) shows a disclosure toggle; open it to see those cards nested directly beneath it. Opening a different source's extractions automatically closes whichever was open — an accordion, so you're never looking at two sources' extractions stacked on top of each other.

Density Toggle

Two icon buttons switch every row between compact (denser, more rows visible at once) and comfortable (larger touch targets, better for mobile). Your choice persists across sessions.

Refine — Filters Inside an Expanded Run (moved in W1.14)

The free-text search and the Well-researched / Verified toggles are no longer top-level rail chrome — they appear only once you expand a run, as a "Refine this run…" control directly above that run's loaded rows. They act only on the rows already loaded for that run:

• Refine this run… (text) — filters the open run's rows by title, query, or body text as you type.
• Well-researched — narrows to Well-documented + Some research sources in this run — useful when writing a report that needs citable, well-detailed references.
• Verified — narrows to sources in this run you've run Verify on, regardless of verdict (Verified / Contradicted / Mixed / No-evidence).
• Clear — appears once a refine filter is active; resets the text and both toggles for that run.

Each run keeps its own refine state — filtering one run's rows doesn't affect any other run, and collapsing/re-expanding a run keeps the filter you set.

Ask-Your-Cafe Scope (Persistent Checkboxes)

Every source row (including "Earlier research" legacy rows — see below) has a checkbox that controls whether that source is included when you ask a question in Ask Your Cafe. Uncheck a source to keep it in your Cafe but leave it out of grounded Q&A answers.

• The rail header shows "N of M in scope" at a glance, with its own toggle-all for the whole Cafe.
• Every expanded run has its own toggle-all control (next to Refine) so you can scope a whole run's sources at once.
• Your choice is saved to your account — it's a persistent setting on the source, not a per-session filter. Come back tomorrow and the same sources are still scoped the way you left them.
• Explicit "ask about THIS source" selections (from the Ask-AI popover) still override the checkboxes for that one question — checkbox scope is the default audience, not an absolute lock.
• In shared/read-only Cafes, viewers see the checkbox state but cannot change it.

If you uncheck every source, Ask Your Cafe still tries to answer from your Serving Tray items — see Filtering and Scoping Sources below for the full behaviour.

"Used in Reports" — Passages from Your Generated Reports

Open any source in the reading pane and look for a "Used in reports" section. This shows the exact passage(s) a generated report cited from that source, tagged with the report that used them. This fixes the "blank reference" problem — a source your report cited is never content-empty; you can always see what the report took from it.

Provenance Link (Clickable ↗)

Every source shows a ↗ (open-in-new-tab) link next to the title, both on the rail row and in the reading pane. Click it to open the original URL — the primary source — in a new tab. This is the canonical provenance link for citations.

Earlier Research

Sources created by the older AI web-search path (before W1.4) — where the AI combined multiple results into a single synthesized response — are preserved exactly as gathered. They do not show a Credibility badge, Verify button, or full reading-pane detail beyond the raw response, and the "Well-researched" refine toggle (see Refine above) hides them.

As of W1.14, these no longer have their own top-level rail section. Since they predate the run/session model, they don't belong to any run — they fall under Individual sources (the same "not produced by a run" bucket as manually-added sources) instead of a dedicated "Earlier research" group. Nothing about the sources themselves changed, only where you find them in the rail.

Legacy rows still get scope checkboxes. Even though their content is read-only, these rows are counted in your "N of M in scope" total and can be checked/unchecked like any other source — since they can still ground Q&A answers, you need full control over whether they're included.

Verify Claim — Live, with auto-produced cited report

Every source has a "Verify claim" button in the reading pane's toolbar. Clicking it opens a dialog where you pick the claims to verify and set a token ceiling, then the loop runs in the background. You can also trigger it by highlighting a passage (or right-clicking a source) and choosing Verify claim from the Ask-AI popover — it verifies the enclosing source's main claim.

• The source resolves to a verdict badge (visible both on the rail row's status dot and in the pane): Verified / Contradicted / Mixed / Couldn't verify.
• The evidence panel in the pane shows the underlying source, supporting evidence, and opposing evidence for each claim.
• A cited report is automatically created in the Reports rail group when the loop completes — one section per claim, with inline [N] citations and a References list containing only the sources cited in the body.

The loop runs in the background — you can close the pane and come back later. See Verify Claim & Cited Report for the full walkthrough.

Find the Truth & Claim Evidence Matrix (W1.11)

Two related features go a step further than Verify claim — instead of checking a claim already sitting in a source, they let you type any claim (or question) and get a probability that it's true, plus a claim-centric view across everything you've checked in the Cafe.

• Find the Truth (every signed-in user, costs tokens) — as of W1.44, reach it from ✦ Deep on the composer, under Interrogate, next to Claims (a typed question in the composer carries into the claim box). Type a claim, the system researches it (and separately hunts for the best evidence against it), then returns a probability (clamped 5–95%, never absolute certainty), a plain-language finding, a credibility-tier-weighted evidence breakdown, the strongest counter-evidence found, and a cited report. Runs in the background and survives closing the tab, with a re-openable "Past truth checks" list.
• Claim Evidence Matrix (every signed-in user, free to view) — as of W1.44, the Claims row (also under ✦ Deep → Interrogate) opens /research/[cafeId]/claims, a table of every claim checked in the Cafe (by Find the Truth or Verify claim) with supporting / contradicting / silent sources grouped by credibility tier and deduped to independent roots (so ten articles quoting one wire story count as one root, not ten).

See Find the Truth & Claim Evidence Matrix for the full walkthrough, including how to read the probability, the "Overall reading" anchor row, and why a Wikipedia-only basis can't push a claim's score much past the mid-80s.

On-Demand Full-Body Fetch

On any source card, a "Fetch full article" action pulls the entire body of the page on demand and stores it in the card for in-column reading. This is optional and gated (it costs a URL scrape). A fetch state indicator shows:

• Not fetched (default) — only the excerpt/abstract is stored
• Fetched — full body available
• Failed — scrape was blocked (e.g. paywall) — the excerpt and metadata are preserved; the card does not regress to a bare reference

---

Source Download

Click the download icon in the reading pane's toolbar for a dropdown with four formats:

• Download as .txt — structured plain text (title, authors, date, journal, DOI, URL, summary)
• Download as .json — full enrichment data as JSON
• Download as .md — the same structured content as Markdown
• Download as .pdf — report-grade formatting (W62): styled headings, bullets, and real tables (not raw | … | pipe text), rendered on the same book theme (SourceSerif4 serif) the report and book PDFs use

PDF, in particular: the title is the card's real display title — the same one shown on-screen, e.g. "Key Points: The Mitochondria Doctor: …" — never a bare "Untitled Source." For an Auto-Research extraction card whose content is really a summary of a transcript or scraped page, the PDF adds a source-attribution line naming the true origin, e.g. "YouTube transcript — <video title> · <url>" or "Web page — <url>" — even for older cards that predate explicit source linking, as long as the parent transcript/scrape is still in the same Cafe.

The filename is a slug of the source's real title (e.g. key-points-the-mitochondria-doctor.pdf) for all four formats, falling back to an identifier (DOI/PMID/arXiv/ISBN) or the Cafe name + a short source ID only when no real title is available. All four formats carry the full body content (title, authors, date, journal, DOI, URL, and summary/body) — not just metadata.

---

Provider Picker (every signed-in user)

The Provider Picker lets you choose which search tools are active for a Cafe. Click the Providers chip row to expand the picker. Provider toggles here control the manual search bar and Auto-Write; the Auto-Research setup has its own per-run tool toggles that work alongside these settings.

Free providers (available to every signed-in user, always on):

• Web Search — AI-powered web search
• Academic Search — OpenAlex, arXiv, PubMed

Apify Actors (every signed-in user, token-billed — each requires a one-time consent per Cafe):

• Google Scholar — Academic paper search via the Google Scholar actor
• Twitter / X — Recent posts on a topic
• Reddit — Community discussions and threads
• LinkedIn — Professional profiles and posts
• Instagram — Posts and profiles
• TikTok — Video posts
• YouTube Search — Video results (distinct from the YouTube Transcript extractor)

Enabling an Apify Actor

Toggle any Apify Actor chip. If you haven't consented to that Actor for this Cafe, a consent modal opens first. The modal shows:

• The platform's terms of service implications (e.g., LinkedIn's scraping policy)
• The approximate per-result token cost for that Actor
• Consent and Cancel buttons

Consenting is recorded per Cafe per Actor — you only see the modal once per Actor per Cafe. On subsequent uses, the Actor is available directly.

Apify Actors are available to every signed-in user — each one still requires the one-time per-Cafe consent above, and each result costs tokens.

Before any of this: the account-level switch (W1.41). Social platform sources — Instagram, Facebook, TikTok, YouTube, X, LinkedIn, Reddit, Scholar (and smaller platforms such as Telegram, Snapchat and Truth Social) — are off by default for your whole account. The first time you start research that could use them, CoffeeScribe shows you the Social Sources Terms in a scrollable panel with three outcomes: Accept (tick the box, then Accept — records your acceptance and enables social platforms everywhere), "Not now" (declines just for this run — your research still runs, those platforms are left out, and the run record names which ones were skipped), or closing the dialog with Escape, the X, or a backdrop click (backs out of the action entirely — nothing starts, nothing is recorded, nothing is billed). Change your mind any time in Settings → Data & Privacy → Social platform sources, which also shows the terms version and date you accepted. The per-Café consent described below is a second, narrower layer beneath this switch: with the account switch off, no per-Café consent can start a social scrape, and there is no administrator exception.

Per-Actor Cost

Each Apify Actor charges a small per-result token cost (higher than the free academic databases). The consent modal shows the exact cost estimate before you enable an Actor.

---

Research Sources

AI Web Search

Type a search query with the AI Web toggle selected. The AI searches the web, synthesizes findings, and presents a formatted response with inline citation links and a numbered Sources section at the bottom. Cost: a small portion of your allowance per search (varies by model). The per-action estimate is shown before you confirm.

Web search works with all models — not just Perplexity Sonar. The AI receives real-time web results regardless of which model you select.

Academic Papers

Click the Academic dropdown to choose from three academic search sources:

• OpenAlex — Search 250M+ works across all disciplines. Includes citation counts. Free, open-access metadata.
• arXiv — Search preprints in physics, mathematics, computer science, and related fields. Includes direct PDF links.
• PubMed — Search 35M+ biomedical and life sciences articles from the NCBI database.

Which database should I use?

Results show paper title, authors, year, citation count (where available), and abstract. Each paper card shows relevant links (DOI, arXiv ID, or PMID). You can:

• Analyze — Send the paper to AI for analysis (Key Findings, Methodology, Compare, ELI5, or custom prompt)
• Add Raw — Save the paper's abstract directly to your Cafe without AI processing
• Dismiss — Remove individual papers from results

The Analyze and Add Raw flows work identically regardless of which academic source you use.

YouTube Transcripts

Paste a YouTube URL and the search bar auto-detects it, showing a "YouTube Transcript" badge. Before extraction, a modal lets you choose:

• Content only — Extract just the transcript
• Summarize, Key Points, Fact Check, Topics & Structure — Extract and analyze in one step
• Custom prompt — Ask anything about the video

You can select multiple options at once.

URL Scraping

Paste any article URL (including www. addresses) and the search bar auto-detects it, showing a "Scrape URL" badge. Same modal options as YouTube — extract content only, or extract and analyze.

Add Multiple Sources at Once (Batch Import)

Pasting two or more sources into the search bar at once triggers the batch import modal automatically — no separate button to find. The bar's placeholder reads "AI web search, YouTube URL, article link — or paste multiple sources at once" as a reminder.

How it works:

1. In the search bar, paste a list of links and/or queries — newline-separated, comma-separated, or just whitespace-separated.
2. As soon as the parser detects 2+ items, the "Add multiple sources at once" modal opens, pre-populated with what you pasted.
3. Each item shows its detected type (Web Search / YouTube / Article / OpenAlex / arXiv / PubMed) with an icon and a per-item token cost. The footer shows the total source count and total cost estimate.
4. Click Run all — items are dispatched to the correct backends with a concurrency cap of 3 (so we don't burst-rate-limit Apify or OpenRouter) and a 1.5s stagger between launches.
5. Each row updates live: Pending → In progress → Done ✓ / Failed ✗. A retry counter appears next to in-progress rows when transient failures (502/503/timeout/Apify/network) trigger an automatic retry.
6. After the batch completes, a summary banner shows how many succeeded, how many failed, how many tokens were charged. Failed items can be retried with one click; sibling successes are not re-run.

Mixed input examples:

Type-hint prefixes (case-insensitive — useful when you want to force a specific backend instead of relying on auto-detect):

• arxiv: pubmed: openalex: academic: — academic backends
• web: search: — force web search (handy when a query happens to look like a domain)
• youtube: yt: — force YouTube (rarely needed; auto-detect handles the common YouTube URL forms)
• url: scrape: — force URL scraping

A line that starts with one of these prefixes is treated as a single item even if it contains spaces (e.g. arxiv: deep RL for robotics is one query, not three).

Batch limit: up to 50 items per batch, for every signed-in user — no account tier involved. If you paste more than the cap, the extras are dropped and an amber "Capped at N items per batch" hint appears. The modal still runs the items it kept.

Failure isolation and retries:

• One bad URL (404, 502, Apify timeout) does not stop sibling items.
• Each item retries up to 3 attempts automatically on transient failures (5xx, 429, timeout, network, Apify-shaped errors). 4xx errors and "no papers found" fail immediately.
• The summary banner is colour-coded: green (all succeeded), amber (partial), red (none succeeded). Failed items can be re-run via the Retry button — successes are skipped.
• The modal blocks accidental dismiss (Escape, click-outside) only while the run is in flight so you don't lose progress mid-batch. After the run completes, you can close it normally.

Cancelling mid-batch: click Cancel. In-flight items are allowed to complete (the server has already done the work and may have already billed for them); pending items are dropped at no charge.

Cost: the same per-item fees as the single-input bar — see the Token Costs table below. The total at the bottom of the modal is the sum of every parsed item's per-type fee, computed before you click Run all.

Analyze Source (on any saved source)

Every saved research source has an Analyze source button (sparkle icon). Click it to open the analysis modal and run AI analysis on any previously saved content — web search results, transcripts, scraped pages, uploaded documents, or imported notebook entries — with presets (Summarize, Key Points, Fact Check, Topics & Structure) or a custom prompt. You can select multiple presets at once.

This button used to be labelled "Ask AI"; it was renamed to avoid colliding with the new Ask-AI popover's Ask AI action, which opens a quick chat about a highlighted snippet rather than running a structured analysis on the whole source. Same capability, clearer name.

The Serving Tray

The right rail is your Serving Tray — a curated collection of the most useful snippets from your research.

Collapsible Rail (W1.13)

The tray defaults to a slim collapsed strip — a count badge plus the top three item titles — so it doesn't eat width you'd rather give to the left panel and reading pane. Click the strip to expand it into the full item list; collapse it again the same way. Your collapse/expand choice is remembered.

Tray Tab vs. Cited Tab (W1.42)

Once expanded, the Serving Tray has two tabs:

• Tray — the snippets and notes you've deliberately saved, exactly as described in this section.
• Cited — different, and automatic: a numbered list of every source actually cited in whatever is open in the centre column right now. Open a chat answer and Cited shows that answer's [1], [2] markers; open a report and Cited shows its reference list. It updates as you move between conversations and reports — you don't add anything to it yourself. Click any row to open that source.

Adding to the Tray

• Highlight text anywhere (a source open in the pane, a Q&A reply, a report) → the Ask-AI popover appears → "Add to tray"
• Add to tray action in the reading pane's toolbar adds the full content of whatever source is currently open
• Import from Notebooks brings in highlights and notes from your Scribes
• Import from another cafe (W1.20) — pull curated research straight from a different Cafe you own; see below.

Import from Another Cafe (W1.20)

Already curated a Serving Tray in one Cafe and starting a new one on a related topic? You don't have to re-gather and re-pay for the same research. Click Add sources → Import from another cafe to reuse it.

How it works:

1. As of W1.42, click Add sources (in the Cafe's view bar) and choose Import from another cafe.
2. Pick one of your own Cafes from the list — the Cafe you're currently in is left out (there's nothing to import from itself). As of W1.43, if you've organized your cafés into groups, the list is organized by group too — a café that belongs to more than one group is listed under each.
3. The picker shows that Cafe's Serving Tray items. Select the ones you want (or Select all) and click Import.
4. Each selected item lands in this Cafe as two things at once: the underlying source(s) it cites are copied into your Sources tab (under Individual sources), and the excerpt + note itself is copied onto this Cafe's Serving Tray, pointing at the new copy.
5. A toast reports what happened — how many sources were added, how many were skipped as duplicates, and the same for tray items.

Copy, never move. The origin Cafe is completely untouched — nothing is removed from it, and its Serving Tray still has everything it had before.

Only your own Cafes are listed. A Cafe someone shared a read-only link with you, or a public Cafe, is not importable — you can only pull from Cafes you own. (Importing from a Cafe shared with you is planned for a future release.)

Duplicates are skipped, not duplicated. If the source a tray item points to is already in this Cafe — matched by DOI, PMID, arXiv ID, ISBN, URL, or title, whichever it has — the import skips inserting a second copy and attaches the excerpt to the copy (or source) already here instead. Re-running the same import a second time doesn't pile up duplicate excerpts either. The result toast always tells you the added-vs-skipped counts honestly, including when everything you selected turned out to already be here.

Imported sources are labelled. A source card that arrived via import shows a small "Imported from ‹Cafe name›" badge, so you can always tell curated-elsewhere research apart from what you gathered directly in this Cafe.

It costs tokens. Each copied source and copied tray item is "embedded" here (indexed for search — see how this works below) so it becomes searchable in Ask Your Cafe — that's a small, one-time embedding cost per item, same as adding any other new source (see Embedding Costs below). You are not charged again for the origin Cafe's original research (the web search, paper fetch, or scrape that found it in the first place) — only for making it usable here.

Available to every signed-in user — no account tier involved. Requires at least one other Cafe with Serving Tray items to import from.

Managing Tray Items

• Drag to reorder items (expand the rail first)
• Click any item to open and read/edit it full-length in the reading pane — items no longer expand/collapse in place in the rail
• Edit notes on any item, from the pane
• Delete with undo (5-second window to restore)

Exporting a Tray Item (W1.25)

Open any tray item in the reading pane and click Download in the pane's footer for a dropdown with four formats: PDF, HTML, Markdown, and TXT.

This works for every kind of tray item — a report-backed item (a snippet you pulled from an Auto-Write, Verify, Find the Truth, People research, or Auto-Research report), a source-backed item (a whole source or a highlighted passage), a saved Q&A answer, a free note you typed yourself, and a combined item (multiple tray items merged into one — see Combining Items below). Before W1.25, most of these kinds had no export path at all; now the same four formats are available regardless of what the item is or how it was created.

Each format carries the item's title, its full text, and — where the item has them — its citations and provenance (which source or report it came from). Every export is free, no token cost, available to every signed-in user.

Source Provenance

Each tray item shows a small label indicating where it came from — for example, "Web: coffee brewing methods" or "YouTube: How espresso works". Items you created manually show "Manual note". This helps you track the origin of each snippet, especially when combining research from multiple sources. The same labels appear in the Workspace Cafe Trays tab and the import modal.

When you import tray items into a Scribe section, the source attribution line includes the original source query and URL (when available), formatted as: Source: {Cafe name} . {query} . {url}.

Source Attribution on Auto-Research Reports (W1.6; storage changed in W1.17)

Auto-Research's synthesised cards — "Key Findings", "Summary", "Best Quotes", "Most Up-to-Date", and "Contradictions & Gaps" — show the sources they were synthesised from as a "Research used" list, the same block every other report type shows (see Reports — Reachable by Run below). Click a listed source to open it in the reading pane. As of W1.17 these are real reports, not Serving Tray items — they no longer land in the tray automatically; find them under the Auto-Research run in the Sources tab, and use the report's own Add to Tray button (in the reading pane) if you want a copy of one in your tray. Manually added tray items and plain notes never had source chips (no synthesis happened).

Saving a Q&A Answer Preserves Its Citations (W1.6)

Grounded Q&A answers work by a technique called RAG (Retrieval-Augmented Generation) — in plain terms, CoffeeScribe finds the passages in your café most relevant to your question and hands them to the AI so it can answer from your own material instead of guessing. When you Save to Tray, Save as source, or Save to Notebooks from one of these answers, the saved item preserves the citations from the answer:

• Save to Tray — the tray item's [N] markers are clickable; click one to open the cited source in the reading pane. The tray item is titled "Cafe chat: {your question}".
• Save as source — the promoted source is saved with the real rag_answer source type (W1.6.1) and grouped under "Saved from Q&A" in the Sources tab. This is a proper source type — not a renamed notebook import. ("rag_answer" is the internal name for "an answer produced by that retrieval technique" — you'll only see it if you inspect the data directly; on screen it just reads "Saved from Q&A".)
• Save to Notebooks — the note is saved as a cafe note (W1.6.1) and grouped under this Cafe's name in Notebooks, separate from Free Notes. Open Notebooks and look for the Cafe's title in the left panel to find it.

In all three cases, the citations remain meaningful and clickable in the Cafe. The promotions do not strip or lose the source references.

Editing Tray Item Text and Title

Click any tray item to open it in the reading pane. A textarea there lets you modify the content directly. Press Ctrl+Enter (or Cmd+Enter on Mac) to save, or Escape to cancel without changes. Each tray item supports a generous character limit per note — long enough for substantial excerpts.

As of W1.42, you can also rename the item. Click the pencil icon next to its title at the top of the reading pane, type a new name, and press Enter (or click away) to save — Escape backs out without changing anything. The new title shows up everywhere that item appears: its card in the tray, and its row wherever the tray lists it.

Adding Notes

Click the + button in the Serving Tray header to create a blank note. This is useful for pasting in content from external sources that aren't covered by the built-in search and scraping tools, or for writing your own observations and annotations alongside your research.

Combining Items

To merge multiple tray items into one:

1. Click the Select toggle in the Serving Tray header to enter selection mode
2. Use the checkboxes to select two or more items you want to combine
3. Click the Combine button that appears
4. The selected items are merged into a single item, with each constituent prefixed by its source label (e.g. [From source: Smith 2022]: ...) and separated by horizontal rules
5. An Undo option appears for 5 seconds in case you want to reverse the merge

The combined item shows a "Combined from N sources" badge so you can tell at a glance it's a merge. Source provenance is preserved — if all constituent items came from the same source, the combined item inherits that source link; if they came from different sources, the combined item is unlocked from any single citation (useful for tray-mode Auto-Write, where each constituent is labelled individually in the prompt).

Combining is useful for assembling related excerpts from different sources into one consolidated reference.

Using Tray Content

Tray items are ready to use in your Scribes, Notebooks, or Workspace. In the Workspace, the Cafe Trays tab lets you browse and paste tray items directly into sections.

Choosing a Model

As of W1.25, there is one model picker for the whole Cafe page, and it lives in the Ask Your Cafe composer — the input area at the bottom of the reading pane, not the search bar. It drives web search, paper analysis, General Ask-AI answers (see General vs. Grounded below), Auto-Research, Auto-Write, Find the Truth, Verify, and Detect Conflicts. Change it once and every one of those surfaces picks up the new choice.

One mode ignores it on purpose: Grounded Ask-Your-Cafe answers always run on a fixed, non-search model regardless of what the picker is set to — see General vs. Grounded for why.

Every signed-in user gets the full picker. Perplexity: Sonar is pinned at the top with a Recommended chip and pre-selected — it searches the web and cites what it finds, which is what most Cafe work needs — followed by the recommended research models, then a collapsed advanced disclosure for the full 350+ catalogue (at your own risk). Nothing here is tier-gated. (The picker in book creation is a different one and still recommends CoffeeScribe Model.)

Each result card shows which model was used, how many tokens were charged, and whether web sources were found. Different models produce different quality results — experiment to find what works best for your research topic.

Research has its own recommended models (W1.28)

The models recommended here are chosen for research, and they are not the same list the rest of CoffeeScribe uses. Book creation wants prose quality; Cafe work wants live retrieval and careful citation. Those pull in different directions, so the two lists are now separate. Changing one never changes the other — if a model is recommended in a Cafe but not when you create a book, that is deliberate.

Venice models (W1.37)

As of W1.37, the Cafe picker also lists models served directly by Venice.ai — a second AI supplier alongside OpenRouter — including DeepSeek V4 Flash 0731 Fast, a faster-serving build of the same DeepSeek weights that no other supplier offers, plus the regular 0731 and Pro builds. They sit in the same picker, carry the same Writing/Research/Cost dots and price labels as every other model, and go through the same rating process — picking one is an ordinary choice, not a special mode.

One thing is different about a Venice model in this release:

• They're research-only. You'll see them in every Cafe picker (Ask Your Cafe, Auto-Research, Auto-Write, Find the Truth, Verify, Detect Conflicts), but not in the book-creation, Quick Read, or Import model pickers.

As of W1.39, Venice models can search the web too — natively, not through OpenRouter's plugin. Pick one for General ask AI and the Web search toggle works like any other model's: an ordinary on/off switch, on by default. Every other capability (writing, analysis, structured research output, and Grounded mode — which is pinned to its own model and never searches, regardless of what's picked) works the same as any other model.

You can also add your own Venice API key in Settings → API Keys — see Using your own API keys (BYOK).

What the Writing, Research and Cost dots mean

Each model shows up to three dot ratings, and they do not all mean the same kind of thing:

• Cost is exact. It is calculated from the model's real current price every time the picker is shown, so it always reflects what you would actually pay. 1 dot is the cheapest in the set, 5 the most expensive.
• Writing and Research are positions within the recommended research set — not absolute quality scores. The strongest model in the set on that axis shows 5 dots, the weakest shows 1. They answer "which of these should I pick", not "how good is this model in the world".

Because those two are relative, adding or removing a model can change the dots on models you didn't touch. A model that was the best of ten can become the second best of eleven without having changed at all. Hover any rating to see what it means, who ranked it, and when.

Where the numbers come from — the models are actually tested, not looked up. Every recommended model is given the same three claims to research, one easy, one medium and one hard, and writes each up as a short cited report. Those reports are then marked:

• Every citation is checked by fetching it. Whether a URL resolves is a measured fact, not an opinion, and dead links count against a model. So does a [1] marker in the text with no source behind it — an answer that looks sourced exactly where a reader would check.
• Relevance is weighed, not just volume. The marker also sees the title of every page cited, so a model that pads its report with twenty-five loosely related links scores worse than one citing eight sources that genuinely address the claim.
• The reports are marked blind. Whoever is marking sees "Model A, Model B" and never a provider name, so a well-known model earns nothing its work did not.

The marking is performed by a language model, disclosed on hover along with the date. That is a real limitation — a model judging other models can favour work resembling its own — and it is shown rather than hidden so you can weigh it.

A model with no dots has not been rated yet. That shows as "not yet reviewed" rather than as an estimate. We would rather show nothing than a number inferred from the model's price.

One thing the dots deliberately do not tell you: whether a model's citations can be trusted. Citation accuracy is the weakest area of every current frontier model, and independent published measurements of it barely exist. Where we have observed a model's citation behaviour ourselves, that observation is written out in plain words in the model's info panel rather than folded into a dot.

Q&A Chat (Ask Your Cafe)

Ask Your Cafe is the reading pane's home view (W1.13) — it's not a floating bubble anymore. As soon as the pane has nothing else open (no source/tray item/report selected), it shows the chat, ready for a question. As of W1.42, the mode chip on the composer reads "Web" or "Cafe sources" and cycles when clicked — those two on-screen labels are the same two modes this doc calls "General ask AI" and "Grounded from cafe sources" (the internal names from before the composer was rebuilt into one row; you'll see both sets of names in this Help Center). The source count, in-scope count, and a "latest report" link that used to sit in a strip above the chat now live in the left panel — the source/scope counts in the Sources tab's footer, and every report in the Reports tab — and the conflicts count moved to the top of the workspace, next to the Cafe's name. On mobile, an "Ask" button opens the same chat. As of W1.32A the chat opens in "Web" mode by default — a web-search-grounded quick chat — with "Cafe sources" as the opt-in for answers drawn strictly from your own research (see General vs. Grounded below).

How It Works

In plain terms: your research is automatically broken into small passages ("chunks") and turned into a searchable index of them ("embeddings") — this whole technique is sometimes called RAG (Retrieval-Augmented Generation), and it all happens locally on the server, so no data leaves it. When you ask a question in "Cafe sources" mode, the system finds the most relevant passages from your in-scope sources and tray items (see scope above), then sends them to the AI along with your question to generate a grounded answer. The default "Web" mode skips all of that — it asks the model directly, usually with live web search.

The chat is ready as soon as you add your first source. While your sources are being processed for search, a spinner indicates that embeddings are in progress — this usually takes a few seconds. You don't need to refresh: the chat clears the spinner and becomes ready on its own the moment processing finishes. You can keep browsing your research in the meantime. If a source can't be processed for search, the chat says so instead of spinning indefinitely — you can still ask questions in General ask AI mode, or re-add the source to retry.

Asking Questions

1. Open a Cafe — if nothing else is open, you're already looking at the chat
2. Type your question and press Enter (or click Send)
3. In the default General ask AI mode, the answer streams in live — your question pins to the top of the view and the answer writes itself in below it (see Streaming Answers). In Grounded from cafe sources mode, the AI answers using only your in-scope Cafe research, citing sources as numbered references like [1], [2]
4. Click any citation badge or reference chip to open what it points to — a café source opens in the reading pane; a web source opens the page

As of W1.17, a citation can point at a whole report, not just a source or tray item. Every report type — Auto-Write, Detect Conflicts, Verify, Find the Truth, People research, and Auto-Research — is embedded into the search index, so a Grounded answer can now cite "Report: {title}" the same way it cites a source. Click that citation and the cited report opens in the reading pane.

General vs. Grounded

As of W1.42, the chat input has one mode chip that reads "Web" or "Cafe sources" and cycles between them when clicked, with a helper line explaining the current mode. (Before the composer was rebuilt into a single row, this was two separate buttons named "General ask AI" and "Grounded from cafe sources" — you'll still see those fuller names elsewhere in this Help Center; they mean the same two modes.) As of W1.32A, "Web" is the default; "Cafe sources" is the opt-in — the default changed deliberately, because the quick "just ask the AI" chat is the front door:

• "Web" (default) — the AI answers from its own knowledge, running on whatever model is currently selected in the model picker (perplexity/sonar by default), usually with live web search — the Web search toggle defaults to on. No Cafe research context is sent — this mode skips embedding/vector search entirely. Answers stream in live as they're written.
• "Cafe sources" — the AI answers strictly from your in-scope Cafe sources, tray items, and any chats in your cafe memory, citing them. Cafe sources is pinned to a fixed model (DeepSeek V4 Flash) and never searches the web — the model picker doesn't reach it at all, on purpose, so pointing the picker at a search-native model for Web can never quietly carry web material into a Cafe-sources answer. If your research doesn't cover the question, it says so rather than guessing.

As of W1.44, your choice is remembered separately per Cafe, not for your browser as a whole — flipping the chip in one Cafe no longer changes what every other Cafe you own opens in. (Before W1.44, the choice was saved per browser, so one flip anywhere made every Cafe open in Cafe sources mode, permanently and invisibly — if that's happened to you, open each affected Cafe once and set the chip the way you want it there; it now stays put.) One exception: arriving via the highlight Ask-AI popover pre-selects Grounded for that question (its meaning is "explain this thing from my research") — unless you've already picked a mode by hand this session, in which case your live choice is respected. When a Grounded answer comes back as a refusal ("your sources don't cover this…"), the answer includes an "Ask the AI directly" button — click it to instantly re-run the same question in General mode without retyping it.

Web Search in General Mode (W1.25)

General mode has its own Web search toggle (on by default), shown next to the model picker whenever you're in General:

• Off, non-search model selected — the AI answers purely from its own knowledge: no web call, no citations.
• On, non-search model selected — the AI searches the web for the question and cites what it finds.
• A search-native model is selected (any Perplexity Sonar) — the toggle is forced on and disabled (a small lock icon marks it), because that model searches the web on every request whether or not you ask it to. A reason line under the row explains this and names DeepSeek V4 Flash as the model to pick if you want a General answer with no web search.
• A Venice model is selected (e.g. DeepSeek V4 Flash 0731 Fast) — as of W1.39, the toggle is a normal, interactive switch, on by default, exactly like any other non-Sonar model's. Venice runs its own live web search directly — it doesn't go through OpenRouter's search plugin — and returns its own citations, so turning the toggle off still gives you an honest no-search answer from the same model.

The point is honesty about where an answer came from, not just the mechanics of the toggle. Whatever the toggle actually did is exactly what the helper line above the composer says happened — never "off" while a search ran anyway. When a General answer cites the web, those citations render as amber chips explicitly labelled "The AI's own web sources — not your café research" — a different color and a different label from the blue chips a Grounded answer's café citations use, so you're never left guessing which pool a reference came from. Click an amber chip to open the page it points to.

As of W1.39, web-grounded General answers cite with the same clickable numbered [1], [2] buttons used in Grounded mode — click one to open the source page it points to. This applies whether the model is a Venice model or an OpenRouter model, so the two providers cite consistently on the same screen; a link the model writes to a page that isn't already a numbered reference stays a plain link rather than inventing a number. The chip row below the answer is the answer's one and only source list: the trailing "Sources" list the model itself writes is hidden everywhere it could otherwise appear — the live answer, the saved answer, scrolled-back history, and anything you copy or save to your Tray or Notebooks — because it would just duplicate the chips.

This costs tokens. A General answer that searches the web is a paid provider call, same as any other AI Web Search action — see Token Costs below. As of W1.39, a Venice search answer is billed from what Venice actually charges us for that call (typically a fraction of a cent) plus Coffeescribe's standard margin, rather than a flat per-search fee — so the token cost you see reflects the real cost of that specific answer. If you're using your own Venice key (BYOK), you aren't charged Coffeescribe tokens for the call at all.

Thinking Toggle (W1.35)

Some AI models can "think" — reason through a question with extra deliberation steps — before writing their answer, and some of those models can be told to skip that step. Whenever the currently-answering model supports both, a Thinking switch appears in the composer right next to Web search.

• Where it shows up. The toggle appears in both General and Grounded mode, and for shared-link viewers too — it's driven by whichever model will actually answer the question (the model picker's selection in General; the fixed Grounded model in Grounded, same as above).
• Only for models that can meaningfully be switched. It renders only for a model that thinks by default and can be told not to. Models that never think at all, models that always think no matter what you ask, and models that only think when explicitly told to (a class deliberately excluded — asking them to think leaks their raw reasoning into the visible answer with no clean way to strip it out) all show no toggle rather than a switch that can't honestly do anything.
• Default off. Off is the model's fast, default behaviour. Turning Thinking on makes the model reason before answering — noticeably slower (roughly 30 seconds up to a couple of minutes, depending on the model and question) and it costs more tokens, since the reasoning itself is billed.
• Remembered per Cafe. Your last choice is saved separately for each Cafe — the same per-café pattern the Web search toggle uses — so you can leave Thinking on in a research-heavy Cafe while your other Cafes stay fast.
• Every answer says how it actually ran. A small 🧠 "thought" marker appears on an answer's meta line (next to the model, tokens, and time) whenever that specific answer reasoned before writing — including after a page reload, when the answer loads back from history. An answer with Thinking off, or one from before this feature existed, shows nothing extra there.
• Shared-link viewers get the same toggle. Anyone reading your shared Cafe sees and can use Thinking too; their questions (and any extra tokens Thinking uses) bill their own account, not yours, and nothing about a viewer's answer is saved to your Cafe's history — same as Q&A on Shared Cafes already worked.

Streaming Answers, Pinned Questions, and the Answer Timer (W1.32A)

General answers stream in live. The moment you send a question, your question bubble appears and pins to the top of the view, and the answer writes itself in underneath — markdown (headings, bold, lists) renders as it arrives, so the text never "snaps" from raw symbols into styled prose at the end. The view deliberately doesn't auto-scroll while the answer grows: your question and the start of the answer stay put, and you scroll down at your own pace if the answer runs long. When the stream completes, the finished answer card takes its place — with its citation chips, model and token labels — without yanking the viewport anywhere.

The wait indicator tells you what's actually happening. Before the first words arrive: with web search on it says "Searching the web…" (that wait is the search), without it "Asking the AI…", and in Grounded mode "Searching your research…".

Every answer shows how long it took. The meta line under each answer — model, tokens, time ago — now also shows the wall-clock answer time, e.g. sonar · 1,204 tokens · 27s · just now. That's the full question-to-finished-answer duration, in both modes. As of W1.35, the same meta line also carries a small 🧠 "thought" marker when that specific answer reasoned before writing — see Thinking Toggle above.

Filtering and Scoping Sources

Which sources ground a Grounded answer is controlled by the checkboxes on the sources rail, not by anything inside the chat panel itself (there's no separate in-chat filter icon anymore — scope lives on the rail so it's visible and persistent, not a session-only setting buried in the chat). See Ask-Your-Cafe Scope above for how to check/uncheck sources and groups.

A quirk worth knowing: your Serving Tray is never scope-gated. Even if you uncheck every single source, Ask Your Cafe still answers from whatever is in your Serving Tray — with an honest "Answering from 0 of N sources" banner on the reply so you know exactly what grounded the answer. This is deliberate: tray items are things you've already curated as worth keeping, so they're always available to Q&A regardless of the source checkboxes. If there's genuinely nothing in scope anywhere (no checked sources AND an empty tray), you'll see a friendly "all sources are toggled off" message instead of a fabricated answer.

Explicit "ask about THIS source" requests (from the Ask-AI popover on a specific source) still bypass the checkboxes entirely for that one question, exactly as before.

Conversation Context (W1.35)

As of W1.35, every chat is a saved conversation, listed in the Chats tab of the left panel (as of W1.42, one of three tabs — Chats · Reports · Sources) — not a scratch buffer that vanishes the moment you move on.

• Starting a new chat doesn't delete anything. Click New in the composer and the conversation you were in is archived — it moves into the Chats list, titled from its first question, with every turn preserved — while the panel clears for a fresh conversation.
• The AI's memory is scoped to the chat you're in, and it's durable (deepened in W1.35 P3). Follow-up questions ("Tell me more about point 3", "How does that compare to the other study?") can reference earlier turns in the same chat — up to the last 20 exchanges, or a size ceiling on very long turns, whichever comes first — but never turns from a different saved chat. The AI's context is rebuilt on the server from the chat's saved turns, not held in your browser tab, so it survives a page reload and reopening a saved chat later — a follow-up question answers with the same memory it had before, not a blank slate. Switch to another chat and the AI still has no memory of what you asked in the one you left; that's deliberate, so an old conversation about one topic can never leak into a new one about something else.
• A retrieval-enabled chat never cites its own transcript back at itself (W1.35 P3). If you've added the chat you're currently in to your cafe memory (see below), that chat's own conversation is left out of what a Cafe-sources answer in it retrieves — so it can't quote its own earlier reply as if it were a separate research source. Other chats you've added stay fully citable from elsewhere; this only affects a chat's answers about itself.
• A stalled answer is retried automatically, not left as an error (W1.35 P3). On the rare occasion a General answer starts writing and then goes quiet, Coffeescribe discards the partial text and asks the question again rather than leaving you with a failure. You may briefly see an answer disappear and a fresh one begin — that's the retry, not a bug — and you're only ever charged once for the answer you receive.
• Delete this chat (the trash icon beside New) permanently removes the turns in the chat you're currently in, behind a confirmation. It does not touch any other saved chat — to remove a different chat, delete it from its own row in the Chats section instead.
• Cafe memory (renamed from "Use in answers" as of W1.42). Every saved chat has a toggle (Sparkles icon, off by default) that adds it to your cafe memory — from then on, a Cafe sources answer can cite that conversation the same way it cites a source you gathered. Flip it on and an "in memory" pill appears on the chat's row, and the button reads "Remove from your cafe memory". Flip it off and the conversation is no longer citable; the chat itself isn't touched either way. Cafe memory only affects Cafe sources answers — it does nothing in Web mode, which never looks at your café's own material.
• Chats from before this feature existed were automatically saved as one archived chat per Cafe when this shipped — nothing was lost, and it's ready to reopen from the Chats tab like any other saved chat.

Suggested Follow-Up Questions (W1.42)

After each answer, up to six clickable follow-up-question cards appear underneath it — click one and it's asked as your next question, no retyping. This is a genuinely new capability, not a rename of anything, so a few things are worth knowing plainly:

• It costs a few tokens per answer, whether or not you click a card. Generating the suggestions is itself a small AI call, made automatically the moment your answer finishes — it isn't optional the way clicking a card is. The cost is small: the call always runs on a fast, cheap model chosen by Coffeescribe, never the model you picked for your chat, so the extra charge per answer is a small fraction of what the answer itself cost.
• BYOK-aware. If you're using your own provider key for that call, you aren't charged Coffeescribe tokens for it, same as everywhere else BYOK applies.
• Turn it off entirely. Open the model chip in the composer and switch off "Suggest follow-ups" (on by default) — no more cards, and no more charge for generating them.
• Hide them for one answer. Click "hide" next to the "FOLLOW UP" heading on any answer to dismiss its cards without turning the feature off.
• They persist. Reload the page or reopen a saved chat and the cards (and any you'd hidden) come back exactly as you left them.
• They read your recent conversation, not just the last answer. The suggestions are generated from up to three earlier turns of the same chat plus the question just asked and the answer just given, so they can build on where the conversation is actually going.
• Not available to a shared-link viewer — their answers are never saved to your Cafe, and there's nothing for a per-café follow-ups preference to attach to.

Chat Attachments (W1.35 P4)

As of W1.35 P4, the composer's paperclip icon — next to the question box, in both General and Grounded mode — is the way to bring a file into a Cafe. Click it for two options:

• Upload a file — one entry covers everything: documents (PDF, DOCX, TXT, MD, EPUB, .html/.htm), images, audio, and video. There's no per-type picker to navigate — pick the file and Coffeescribe classifies it automatically. Size and length limits: documents up to 25 MB, images up to 10 MB, audio up to 15 minutes, video up to 3 minutes (a duration check, not a byte cap — CoffeeScribe shrinks the file before sending).
• From café sources — pull in a source already gathered in this Cafe instead of uploading a new file. Only shown in General mode (see below for why).

Every upload becomes an ordinary café source — rail-visible, openable, and previewable, exactly like a source you searched or scraped in. The original file itself is owner-only; nobody else can download it, even from a shared link.

What happens next depends on the mode you're in when you attach it:

• General mode — the file also stays with the chat. It's listed under that conversation in the Chats section of the rail with a small paperclip count, and its full text is injected into every later question in that same conversation — so "Summarise this," "What's this about?," or a specific follow-up all work without re-attaching anything. The most recently attached file(s) are called out to the model as what "this" most likely refers to; older attachments in the same conversation are noted as attached earlier, so a five-file conversation doesn't get confused about which file a follow-up means.
• Grounded mode — the file becomes a café source and stops there. It reaches a Grounded answer, if at all, through the same ranked retrieval as every other in-scope source — nothing is force-injected, and it isn't bound to the chat (this is also why "From café sources" isn't offered in Grounded — there's no chat binding for it to do).

Media files are converted to text automatically, at upload time. Images, audio, and video are described/transcribed by an AI model as soon as they finish uploading — not re-processed on every question — so a 15-minute audio clip is transcribed once and every later question reuses that transcript. This is a billed step (BYOK-aware, like any other AI call), and for a large upload the estimated cost is shown before it's charged. If the conversion comes back empty (a rare provider hiccup), Coffeescribe retries once rather than silently saving a blank, unanswerable source.

Send is blocked while a file is still uploading or being converted. The Send button disables with a "Your file is still uploading" note until the attachment finishes and appears as a chip — this prevents a question firing before its own file is actually available to answer from.

A chat nearing its model's context limit warns you, then blocks you. As a General conversation (with its attachments and history) approaches the current model's context window, an amber notice appears: "This chat is getting close to the model's context limit. Start a new chat, or switch to a model with a larger context window." If you keep going, a hard limit follows: sends are blocked outright until you start a new chat — your typed question is kept in the box, not lost.

A large attached document roughly doubles the cost of an answer, because the whole document is read again on every question in that chat (not just the one where you attached it) — this is expected, not a bug, and is the trade-off for the model actually being able to answer from the whole document rather than a partial excerpt.

Deleting a chat asks what should happen to its uploads. If the chat you're deleting carries files, a choice appears: keep them (default — they stay as ordinary café sources you can still search and read) or delete them too (removes the files uploaded through that chat). Either way, a source you merely picked from café sources into the chat (rather than uploaded through it) is never deleted by a chat delete — only unlinked.

Shared-link viewers get none of this. There's no attach control, no pending-upload chips, and no Chats section on a shared Cafe — a viewer's questions never carry an attachment. Attachment text you uploaded is shared as an ordinary source like any other (viewable, citable), but the original file itself stays owner-only.

Actions on Answers

Each answer has action buttons:

• Copy — Copies the answer with a full reference list appended
• +Tray — Save the answer to your Serving Tray. The tray item is titled "Cafe chat: {your question}" and its [N] markers are clickable — click one to open the cited source in the reading pane. Citations are preserved.
• +Source — Save as a new research source. The promoted source is grouped under "Saved from Q&A" in the Sources rail.
• +Notebook — Save to your Notebooks as a free note with citations preserved. Find it under Free Notes in Notebooks.
• Delete — Remove a single Q&A entry (just that question and answer — to remove a whole conversation, use Delete this chat or the delete control on the chat's row in Chats)

Editable Source Titles

Double-click any research source title to rename it. This helps you organise your sources when the auto-generated titles aren't descriptive enough. Tray items can also have custom titles.

Embedding Costs

Each research source and tray item is automatically chunked and embedded when created. This is a small one-time cost per item, broken into chunks of a few thousand characters each. The per-action estimate is always shown before you confirm.

Model

General answers (the default mode) run on whatever model the picker is currently set to (perplexity/sonar by default). Grounded answers are pinned to DeepSeek V4 Flash, a model chosen for its ability to follow strict instructions and answer only from the context it's given — regardless of what's selected in the model picker. See General vs. Grounded above for the full picture, including the web-search toggle and how a General answer's web references are labelled.

Sharing a Cafe

You can share any Cafe with other Coffeescribe users via a read-only link.

Enabling Sharing

1. Open a Cafe and click Share in the view bar (top of the workspace)
2. Toggle Enable sharing on
3. A share URL appears — click Copy to copy it to your clipboard
4. Send the link to anyone with a Coffeescribe account

What Viewers Can Do

Viewers who open your shared link can:

• Browse all your research sources and their content
• Read your Serving Tray items
• Ask Q&A questions about your research (using their own tokens, not yours)
• Copy text from sources and tray items

What Viewers Cannot Do

Viewers cannot search, add sources, modify tray items, delete anything, save Q&A answers to your Cafe, or attach a file (W1.35 P4) — the composer's attach control is not rendered at all on a shared link. All edit controls are hidden. Read-only access is enforced at the database level.

Q&A on Shared Cafes

Viewers can ask Q&A questions in your shared Cafe. Their questions use their token balance (not yours), and their Q&A history is session-only — it is not saved to your Cafe and disappears when they close the page. As of W1.35, viewers also see the Thinking toggle for models that support it — if a viewer turns it on, the extra reasoning tokens bill their account too, not yours. Chats and saved conversations are owner-only (W1.35 P2) — the shared view never shows a Chats section, nothing a viewer asks is ever saved as a chat, and there is no "New"/rename/delete/Use-in-answers UI for them.

Revoking Access

Toggle sharing off in the Share dialog to immediately disable the link. Anyone with the old link will be redirected away.

Regenerating the Link

Click Regenerate link to create a new share URL. The old link immediately stops working. A confirmation dialog warns you before regenerating.

Share Indicator

Shared Cafes show a small share icon next to the cafe name on the Research listing page, so you can easily see which Cafes are currently shared.

Groups Are Never Shared (W1.43)

If you've organized cafés into groups, that organization is yours alone — a shared read-only link never reveals which group(s) a café belongs to, or that you have groups at all. Group membership is invisible on every shared-café surface, with no exception.

Auto-Research (W1.5b)

Auto-Research is a steerable AI agent that researches a topic for you across a wide range of tools. You control what it uses, how deep it goes, and what it extracts — the agent turns those controls into a plan, executes it step by step, and sends the results into your Sources rail and Serving Tray.

As of W1.44, click ✦ Deep on your Cafe's composer, then Auto-Research under Gather — or a depth chip under its row for a one-tap start at that depth — to open the setup in the centre column. A topic already typed in the composer comes with you as the starting topic.

Quick Mode — A True Fast Answer (W1.32B)

Auto-Research now has a true fast mode, and it's the setup's default: Quick is the first option in Research Depth, above Solid research and Leave no stone unturned (renamed labels for what were previously called Standard and Deep). Ask it anything and it plans its own research in seconds: it picks the best specialist gatherers for your question — social platforms, academic papers, news archives, the live web — runs them all at once, and hands you a fully cited report a few minutes later, with every source saved into your café to keep, question, and build on. One box, one click; the depth dial is there when you want to go further.

Under the hood, Quick runs on a different, faster engine from Solid/Deep (described in the rest of this section) — not just the same agent with a lower ceiling:

• Plans in one quick pass. A single planning call reads your topic and picks up to 7 gatherers — drawn from a wide shelf of web/social scraping tools (the same kind of tools behind the consent-gated tools listed below) plus this product's own research tools (OpenAlex, PubMed, Wikipedia, newspaper archives, and more) — and a live web search. Nothing is chosen by you; the plan is generated for you, the same way "Let the agent choose tools" works for Solid/Deep (see Tools below).
• Everything runs in parallel. Every gatherer and the web search fire together rather than one after another, which is most of why Quick is fast.
• Every result becomes a real source, saved to your Sources rail exactly like any other Auto-Research run — nothing about how sources are stored or cited is different for Quick.
• A Quick run always writes a full cited report. Unlike Solid/Deep (see Full Report below), there's no "sources only" option for Quick — see the note there for why.
• Extraction cards work the same as always — see What to EXTRACT below; if you leave the pills untouched, a sensible set is picked for you.
• When it finishes, it offers to go deeper. See Go Deeper? below.
• The setup is simpler in Quick. There's no Advanced panel — no manual tool picker, no results-per-search control, no "research only" toggle, no prompt/skills/fan-out editing. Those exist for Solid/Deep, described later in this section; Quick is deliberately just a topic box and a Start button.

Go Deeper? (W1.32B)

Once a Quick run finishes, a "Go deeper?" panel appears with twelve concrete next moves — two rows of six — built from what your run actually found (it names real sources and angles from your results, not generic suggestions).

• Picking one pre-fills a new Auto-Research run at Solid or Deep depth, with the topic already filled in for you. Nothing launches and nothing is charged just by clicking an option — it opens the setup ready to go; you still click Start to actually run it.
• It's generated once, right when your Quick run finishes, from that run's own topic and sources — not regenerated every time you look at the completed run.
• No offer, no charge for it. If a run gathered nothing worth building suggestions from, no "Go deeper?" panel appears at all, and nothing is charged for the attempt.

Runs in the Background — Survives Closing the Tab

Auto-Research runs as a server-side background job. Once you click Start Research, the run continues on the server regardless of what you do in your browser:

• Close the tab, refresh, or navigate away — the run keeps going. Come back to the Cafe and your sources will be there when it finishes.
• Reattach after a refresh — if a run is still in progress when you reload the page, the live progress panel comes back automatically.
• Sources land as the agent works — sources are saved to your Cafe as each step completes. Reports (your selected extract types) are added once the full run finishes.

Note: Reports are only produced when the run completes. If you cancel mid-run, sources gathered before you stopped are kept, but no reports are produced — the synthesis step only runs on completion.

Waiting for a Free Slot

Two caps govern deep background runs — Auto-Research, Find the Truth, and Research a Person — and either can make one wait: your own cap on how many of your runs execute at the same time (the one you'll meet most often — start an Auto-Research while a person research is still going and the second waits), and the platform cap across all users, so a busy moment degrades to waiting rather than to every run slowing down.

If either cap is full when you click Start Research, a "Waiting for a free slot" panel appears in place of the usual progress panel — it shows your position in line (e.g. "You're #3 in the queue"), starts your run automatically the moment a slot frees up, and makes clear that nothing is charged while you wait. You can close the window and the run holds its place; a Cancel button lets you give up your spot. The wait is bounded: if no slot frees up in time the run is released rather than hanging forever, nothing is charged, and you can start it again. The release message says which cap held it up, because they need different responses — "You already have a run going" means finish or cancel that one, while "The platform is at capacity" or "Other runs were ahead of yours" means try again shortly. Both caps apply identically to BYOK runs — they protect shared database capacity, not billing.

---

What to EXTRACT

Choose what the agent synthesizes into a report at the end of the run — available on every depth, Quick included. Select one or more from eleven presets (grown from five as of W1.32B, each with its own distinct instruction and length so picking several gets you genuinely different cards, not near-duplicates):

• Key Findings — the most important takeaways from the research
• Summary — a concise overview of what was found
• Plain-English Explainer (W1.32B) — the topic explained to someone new to it, jargon-free
• Best Quotes — the most citable or representative passages
• Most Up-to-Date (labelled "Latest Developments") — the most recent information on the topic
• Contradictions & Gaps — conflicting claims and missing evidence
• Timeline (W1.32B) — how the topic developed over time
• Statistics & Data (W1.32B) — the concrete numbers the sources report
• Expert Opinions (W1.32B) — what named authorities/specialists said
• Pros & Cons (W1.32B) — the case for and against
• Action Items (W1.32B) — practical next steps a reader could take

Picked for you if left empty. You don't have to tick anything — leave the pills alone and a sensible default set is chosen automatically, the same set a Quick run pre-ticks when the setup opens. Want something none of the eleven cover? Describe it in the "Want a different extraction?" box right below the pills.

Each becomes its own report card under this run in the Sources rail's Auto-Research section when the run completes (W1.17 — these are full reports, not Serving Tray items; see Reports — Reachable by Run below for the full action set). Use the report's own Add to Tray button if you want a copy in your Serving Tray.

---

Full Report (W1.18)

Right below the EXTRACT presets is a second, larger box: Full Report. This is a different thing from the extraction cards above — instead of several short synthesized cards, it's one complete written report on your topic, with citations, written by the same engine that powers Auto-Write.

• On by default for every depth, as of W1.32B (it used to be off by default). This box is a status display, not a click-to-toggle switch — what you can do with it differs by depth:
  • Quick always writes one, with no way to turn it off. A Quick run's whole point is a fast, cited answer, so this row shows "Every Quick run writes one" and there's nothing to click. If you only want raw sources with no synthesis, use Solid or Deep instead and turn the report off there (see the next bullet) — Quick doesn't offer that trade.
  • Solid/Deep default it on too, but you can turn it off — under Advanced → "Research only (no report)". Tick that and this row updates to say "Skipped."
• It reuses the run's own sources — no second gather. When the run finishes gathering and writing its extraction cards, the SAME background job continues and writes the full report using only the sources that run just collected (ranked by quality and capped to your chosen depth). It does not search for new sources, and Auto-Write itself is untouched.
• It costs extra tokens. Writing the report is an additional pass on top of the research itself, so the cost estimate shows a second line: "Research ~X + Report ~Y" — X is the research cost you'd pay anyway, Y is the extra cost of writing the report. Turning it off (Solid/Deep only) means no extra cost. The estimate is shown before you start, and the run won't proceed if you can't afford both.
• It shows up as a "Full report" card. Once the run completes, the report appears under this run in the Sources rail — labelled "Full report" so you can tell it apart from the extraction cards — and opens in the reading pane with the same full action set as any other report: Share, Add to Tray, Delete, Copy, Download .md, Download PDF, and the citation toggle. Like every report type (W1.17), it's also searchable and citable in Ask Your Cafe.
• A report failure doesn't fail the run. Sources and extraction cards are already saved by the time the report starts writing, so if the report-writing step itself fails, your run still completes normally — you'll see a "Report failed" note next to the run instead of a report card, and everything else about the run (sources, extraction cards) is unaffected.

A note on quality: the Full Report can only be as good as the sources the run gathered. As of W1.18 Step 2, every gathered source is automatically tagged with a topical-relevance score, and sources scoring clearly off-topic are excluded from what a report cites (they're still kept, tagged, in your Cafe as raw sources — nothing is silently deleted). This closes most of the drift issue described in the FAQ. A source that's borderline-relevant, or one you explicitly pin into a report yourself, can still slip through — the relevance gate is the primary defense, not a guarantee — so reading the extraction cards first remains a good sanity check before relying on the full report.

---

Tools

The agent can search across a wide range of sources. As of W1.32B, "Let the agent choose tools" is on by default for every depth — the setup shows this as a locked, always-on status row, and the agent picks which of the tools below to use for your question. This is the same underlying idea as Quick Mode above choosing gatherers for you, just for Solid/Deep. To pick tools yourself instead, open Advanced (Solid/Deep only — Quick has no Advanced panel) and toggle any tool below manually; the moment you do, the setup switches out of auto-pick for that run, with a "Reset to auto" link if you change your mind. The tools themselves are unchanged by this — only who decides which ones run:

Always-free tools (no consent required):

Consent-gated tools (every signed-in user, token-billed):

Each of the following requires your account-level acceptance of the Social Sources Terms (W1.41 — see the note above) and a one-time per-Cafe consent the first time you enable it. The per-Café consent modal explains the platform's terms, what data is sent, and the approximate cost per result. After both are in place, the tool is available immediately for that Cafe.

What consent means: These tools use Apify actors to access social platforms. Your search query is sent to Apify; it runs the actor on its own infrastructure. Your queries are not attributed to your personal account on those platforms. Consent is recorded per Cafe per tool — you only see the modal once per tool per Cafe. See the Privacy Policy and Third-Party Data Sources section for the platforms' terms of service links.

Availability: The six consent-gated tools are available to every signed-in user — enabling one requires the account-level social-sources switch to be on and the per-Cafe consent above, and each result costs tokens.

---

Newspaper Archives (W1.12a)

Newspaper archives is an always-free Auto-Research tool (no consent needed, no separate charge) that searches digitised historic newspapers and periodicals — the archival/historic counterpart to the present-day News tool above. It's aimed at questions like "how was this reported at the time," obituaries, old local history, and claims that predate the modern internet.

What it searches: results are pooled from Internet Archive and Europeana (Europeana covers European national/regional newspaper collections). Chronicling America (the Library of Congress's historic-newspaper API) is wired into the tool but is currently blocked by Cloudflare from our servers — Internet Archive and Europeana carry full coverage in the meantime; Chronicling America results will resume automatically once that's resolved, with no action needed from you.

Honest, tiered storage — nothing is ever excluded. Every newspaper hit becomes a real source in your Cafe, but how much of it is stored depends on its copyright/licensing status:

A snippet-only or reference-only source is never hidden or dropped — it's an honestly-labeled, citable source you can always open the original link for.

Feeds Auto-Write and Find the Truth automatically. Newspaper sources carry real bibliographic metadata (publication name, date, byline where available), so Auto-Write renders proper newspaper citations in all eight citation styles, and they're treated as ordinary pool members — no separate step required. In Find the Truth, newspaper evidence is weighed claim-type-aware: a newspaper article is strong evidence for whether something happened or was reported at the time (e.g. "was the 1918 flu widely covered in the press?"), but weak evidence on its own for whether an underlying claim is actually true (e.g. a period newspaper repeating a rumor doesn't make the rumor true). Snippet-only and reference-only sources are capped so they can't single-handedly decide a verdict.

---

Research Depth

Quick runs on genuinely different machinery from the other two — everything from here through Live Progress and Cancel below (tools toggles, Advanced panel, fan-out, the reflect-and-replan loop, the Research Intent Interview) describes Solid research and Leave no stone unturned. Quick has its own, simpler setup — see Quick Mode above for what it does instead.

Important about "Leave no stone unturned": The token budget is a hard ceiling, not a target. The agent uses its own judgment and stops when it believes it has gathered enough, which is often well before the budget is exhausted. "Deep" means the agent is instructed to be thorough — it does not mean it will burn your entire budget. The pre-flight cost estimate is shown before you start; a Deep warning appears in the setup when Deep is selected.

Deep can reflect and replan (new). After a Deep run's first research pass, the agent briefly checks its own work: "did this leave any real gaps?" If it did, it runs up to two more focused rounds chasing the specific unanswered facets or contradictions — you'll see a second (or third) "Research plan ready" step appear in the live progress log, which is the agent digging deeper rather than a glitch. If the first pass already answered the topic well, it stops there — a well-covered topic finishing in one round is the correct, expected outcome, not a shortfall. Either way, this stays inside the same token budget you set (extra rounds don't get a fresh ceiling), and there's a hard cap of two extra rounds regardless of how thorough the agent thinks it could still be. This only applies to Deep — Quick and Standard never reflect or replan.

Deep sets Fan-Out to Auto by default. Because fan-out (decomposing your topic into sub-questions — see Fan-Out below) is the natural companion to a thorough run, selecting Deep sets fan-out to Auto (the planner decides); Quick and Standard set it to Off. This is a default, not a lock — change it to On or Off in Advanced, and your explicit choice sticks even if you change depth afterwards. When Auto is active, the pre-flight estimate shows a rough number with a warning, because the agent decides at runtime how many sub-questions to create.

---

Research Intent Interview (W1.18 Step 2, Deep only)

When you pick Deep depth and click Start Research, a short interview step appears before the run begins: 2–5 multiple-choice questions — a few fixed questions about your intent, plus a few generated for your specific topic — each with an always-available "Something else…" free-text option.

• Always skippable in one action. Click Skip and the run starts exactly as it would today, on the bare topic — skipping is never a worse outcome, just the same run you'd have gotten before this feature existed.
• Answering shapes the run. Your answers compile into a short brief that reaches the agent's planner first (before it drafts a research plan) and again at the synthesis/writing step, so the plan and the final report both stay aligned with what you actually asked for.
• Only appears on Deep runs. Quick and Standard skip straight to research — the interview isn't shown, no skip needed.
• If question-generation is slow or fails, the fixed questions still show and Skip is still available — the interview never blocks or delays your run indefinitely.
• Your brief is saved with the run and feeds the closing ledger below (briefShaped: true/false).

---

Results Per Search (Top-k)

The Results per search control (5 / 10 / 20) sets the candidate pool size — how many results each search tool fetches for the agent to consider. This is separate from depth:

• Depth = how many cycles / how hard the agent tries
• Results per search = how wide a net each individual search casts

Higher top-k gives the agent more to choose from per step but increases per-step cost.

---

Advanced Controls

Solid/Deep only — Quick has no Advanced panel at all (see Quick Mode above). Click Advanced — prompt, skills & loop to expand:

Prompt Sections

The agent's behaviour is driven by five editable prompt sections. Each has a sensible default — edit to shape the run, or click Reset to restore the default:

• Role — what kind of researcher the agent acts as
• Method — how it approaches the topic (broad-to-deep, query variation, source routing)
• Evidence & Skepticism — how hard it tries to disprove claims before accepting them
• Source Hierarchy — how to weight different source types (peer-review > preprint > official > news > social)
• Output Contract — what to extract and how to format it

Note: The JSON-plan contract (the machine-readable format the agent uses to structure its plan) is always appended by the system after your text and cannot be overridden. Editing prompt sections changes the agent's strategy, not its ability to produce a parseable plan.

Research Skills

Toggle preset research playbooks on or off:

• Systematic Literature Review — structured, comprehensive, tracks methodology
• Fact-Check Claims — actively seeks disconfirming evidence
• Find Opposing Views — seeks counter-arguments and alternative perspectives

Each skill injects additional instructions into the agent's planner and synthesis steps.

Fan-Out (Sub-Question Decomposition)

Fan-out controls whether the agent decomposes your topic into parallel sub-questions or researches it as a single linear sweep. The control is now a tri-state:

Deep selects Auto by default — because a thorough run naturally benefits from letting the agent decide how to structure the research. Quick and Standard default to Off. These are only defaults: switch to any setting in Advanced, and your explicit choice sticks even if you later change depth.

When Auto is active, the pre-flight cost estimate shows a rough number — because the agent decides at runtime whether to decompose and how many sub-questions to create. A "Rough estimate — the agent decides how deep to go, up to your cap of X tokens" warning appears near the estimate. The executor's hard ceiling is still honoured: the agent always stops at or before your token budget, and billing charges only actual tokens (the estimate is advisory).

---

Live Progress and Cancel

While the agent runs, a live progress log shows each step (tool, query, step result), running token cost, and source count. Sources appear in the Sources rail as they're found.

Cancel: Click Stop at any time. The server-side job receives the cancel signal and stops at the next step boundary. Sources gathered before you stopped are kept. Reports are not produced for a cancelled run.

Per-step timeout: If a tool step hangs (for example, a slow Apify actor cold-start), it automatically times out after 90 seconds, is marked as skipped, and the run continues. You will never see an infinite spinner.

---

Ambient Conflict Nudge (new)

After any Auto-Research run completes, CoffeeScribe automatically runs a quick, free, lexical-only check across your Cafe's sources for potential factual disagreements. As of W1.42, if it finds any, a calm amber "N conflicts" count appears next to your Cafe's name and source count, at the top of the workspace — click it to open the same Detect Conflicts setup reachable from ✦ Deep (as of W1.44).

• This check is completely free — no AI model call, no tokens charged, no extra step for you. It just looks at word-overlap between source titles/abstracts to flag pairs worth a closer look.
• The badge is a nudge, not a verdict — it's telling you "these sources might disagree," not "these sources definitely disagree." Click the badge to open the same Detect Conflicts setup you'd reach from ✦ Deep — running the actual AI-powered conflict check (which costs tokens, per Detect Conflicts above) is still something you choose to do.
• The count is capped for display (you'll never see an alarming "47 potential conflicts" — the badge tops out at a small number) and is durable — it's saved with the research run, so it's still there if you come back to the Cafe later, not just while the run was live in front of you.
• No badge appears if the check finds nothing, or on a Cafe without enough overlapping sources to compare.

Closing Ledger (W1.18 Step 2)

Solid/Deep only — Quick doesn't produce this ledger (it has no fan-out, no reflect-and-replan, and no Research Intent Interview for the ledger to report on).

When a Solid or Deep run finishes, an honest end-of-run summary — the closing ledger — is attached to it, giving you a transparent look at what the agent actually did rather than a black box:

• Per-sub-question source counts — if the run decomposed your topic (fan-out), how many sources each sub-question actually returned, including any that came back empty.
• Class mix achieved vs. planned — the mix of source types (academic, news, web, social, etc.) the run targeted versus what it actually gathered.
• Leads harvested vs. chased — sources sometimes mention other citable sources or links (leads); the ledger shows how many were spotted and how many the run followed up on.
• Thin-coverage notes — an honest flag on any sub-question that came back with little usable evidence.
• Whether the run was brief-shaped — true if you answered the Research Intent Interview above, false if you skipped it.
• Rounds and replan reasons — for Deep runs that reflected and replanned (see Research Depth above), why each extra round happened.
• Budget-gate stops — whether the run hit its token ceiling before it judged itself done.

The ledger is for transparency, not action — nothing about it changes billing or behaviour after the fact. It's most useful for understanding why a run's report turned out the way it did.

---

Budget Control

Set a token budget before starting. This is a hard ceiling — the agent stops before or at this limit and synthesizes with whatever it has gathered. The pre-flight cost estimate is shown in the setup before you commit.

Auto-Research is available to every signed-in user. The CoffeeScribe Model is pre-selected by default, and every user can choose a different model from the picker.

---

Third-Party Data Sources

The consent-gated social and scholar tools access content via Apify actors. By enabling them you acknowledge that:

• The data is accessed via third-party actors and is subject to each platform's terms of service.
• Content from these platforms is for use in your own research within those terms.
• Apify is the access layer; see the Apify Privacy Policy.

Relevant platform terms:

Legal note: An AUP / indemnification clause for automated access to third-party platforms is a legal-team follow-up item and is not yet included. Use these tools responsibly and within each platform's terms.

---

Availability

Research a Person (W55 / W55.1)

Research a Person turns a name into a compiled, cited dossier instead of checking one claim. As of W1.44, reach it from ✦ Deep on the composer, under Gather (it goes out and brings sources back, the same job as Auto-Research with a person as the subject — it does not sit next to Find the Truth, which checks material you already have). Two stages: a free identity confirmation (pick the real person from a Wikidata/Wikipedia candidate list, or continue with just your own details if nobody matches), then a paid, footprint-gated background sweep (10–40 minutes) that only unlocks social-platform research once CoffeeScribe re-confirms the person is a real public figure. Its setup always opens fresh — unlike Auto-Research and Auto-Write, it doesn't preserve a half-filled draft between visits (see Starting a Run above).

The confirm step lets you tick which research lanes to run (life story, social, own words, perception, persona-specific lanes), add up to 2 custom focuses in your own words (e.g. "his death"), and pick a depth — with a live cost estimate that updates as you adjust any of those. The report leads with an era-by-era life story, quotes the actual text of a public figure's most-liked and most notable social posts with real engagement numbers, and includes a "Who They Are" synthesis plus whichever persona sections apply (Works & Reception, Wealth & Business, Political Views, Death & Aftermath, and more).

Hard guardrails, always on: no contact information or addresses are ever compiled (checked at two separate points in the pipeline); a non-public or unconfirmed person's run stays clean-core-only with an honest limited-coverage note; a deceased person's still-active accounts are labelled estate-run and are never quoted as their own words; an unverified identity match is always framed as "possibly related, unverified," never asserted as fact.

See Research a Person for the full walkthrough — Stage 1 candidate picker, the lane/focus/depth confirm step, what's in the report, and every guardrail.

---

Other Features

Notebook Import

As of W1.42, click Add sources (in the Cafe's view bar) and choose Import from Notebooks to bring in highlights and notes from any of your Scribes. Choose to import as Research Material (left panel) or directly to the Tray (right panel).

Document Upload — Now via the Chat Composer (W1.35 P4)

As of W1.35 P4, uploading a document from Add sources is retired. Bringing a file into a Cafe now happens from the paperclip icon in the Ask Your Cafe composer — the same control that starts every chat — not from Add sources. See Chat Attachments below for file types, size limits, and how an upload interacts with Web vs. Cafe-sources mode. Paste text, Import from Notebooks, and Import from another cafe are unaffected and still live behind Add sources (as of W1.42, a button in the Cafe's view bar, replacing the retired workspace toolbar's "+ Add sources" menu).

Deleting Sources

Hover over any research source to see the delete button. Deletion includes a 5-second undo window.

Deleting a Whole Run (W1.14)

Every run row in the sources rail — Auto-Research, Auto-Write, Detect Conflicts, Find the Truth, People research, Verify — has a trash icon. Click it to open a confirm dialog; confirming deletes that run's report/verdict and its sources from the Cafe in one action, instead of removing sources one at a time.

Anything you've already saved to your Serving Tray is kept. Deleting a run only removes what's still living in the sources rail under that run — a tray item you pulled from it earlier is a separate, deliberately-saved copy and is never touched by a run delete.

This is a hard delete behind a confirm step, not an undo window like the single-source delete above — the confirm dialog is the safety net, so read it before confirming. Individual sources isn't a run and has no whole-group delete — remove those sources one at a time as described above.

Token Costs

Every action below is available to every signed-in user — availability doesn't vary by account. Cost is deducted from your token balance regardless of account type.

The per-action estimate is always shown before you confirm so you know what you're spending.

Availability

Every feature below is available to every signed-in user — none of it is gated by account tier. Token balance is the only limit; social-platform tools additionally require the account-level Social Sources acceptance (W1.41) and the one-time per-Cafe consent described above.

Reports — Reachable by Run (W1.6.1; restructured in W1.14; joined by Auto-Research in W1.17)

As of W1.14, report types are no longer one unified "Reports" group — each report-producing run kind gets its own top-level section in the sources rail:

• Auto-Research — the extraction cards ("Comprehensive Summary", "Key Findings", "Most Recent", etc.) synthesised at the end of a run (W1.17 — see above), plus, if you ticked Full Report (W1.18), a single "Full report" card written over that run's own sources — see Full Report above
• Auto-Write — prose research reports generated by the Auto-Write button; each report is its own run
• Detect Conflicts — factual disagreements detected by Detect Conflicts
• Find the Truth — probability-verdict runs, each with a cited report
• People research — dynamic dossiers produced by Research a Person (W55 / W55.1); each run is labelled with the person's name
• Verify reports — the cited report automatically produced whenever a Verify claim loop completes

Runs appear newest-first within their section. Every run row shows an "open output" button (file icon) that jumps straight to its report in the reading pane — you don't need to expand the run first, and reports no longer expand inline in the rail. This means a re-verified source never orphans its prior report — every Verify run stays listed in the Verify reports section, so old reports stay reachable even as a new one is added.

Every report type is now a real report with the full action set (W1.17). Auto-Research's cards used to be the one exception — they were stored as Serving Tray items, so they opened with only Copy/Edit and could not be shared, PDF'd, or added to the tray. As of W1.17 they're stored the same way every other report is, so Open full, Share, Add to Tray, Delete, Copy, Download .md, Download PDF, and the citation toggle all work identically across Auto-Write, Detect Conflicts, Verify, Find the Truth, People research, and Auto-Research.

Reports still auto-drop into the Serving Tray for Auto-Write and Verify only. When an Auto-Write or Verify report is created, a tray item is automatically added, same as before. Click that tray item and its reading-pane view shows an "Open report" link at the top so you can jump straight to the full-screen artefact viewer without hunting through the rail. Auto-Research, Detect Conflicts, Find the Truth, and People research reports do not auto-drop — the Serving Tray stays a purely manual shelf for those (W1.14); use the report's own Add to Tray button if you want a copy there.

Deleting a run removes its report too — see Deleting a Whole Run above.

Every report type can be downloaded as a book-grade PDF (W54). Alongside Copy and Download (.md), the report viewer's Download PDF button saves a single, typeset A4 PDF for any Auto-Research, Auto-Write, Verify, Conflict, Find the Truth, or person research (W55) report — the Find the Truth PDF additionally includes the verdict graphic and claim/truth matrix in a fixed order. Generated entirely by code, free for every signed-in user, no token cost, and available to anyone a report is shared with. See Find the Truth & Claim Evidence Matrix for the fixed section order.

Save the whole report to Notebooks (W1.17). Every report type has a Save to Notebooks button (beside Copy / Download .md / Download PDF) that saves the entire report body as a note grouped under this Cafe's name in Notebooks — separate from the older selection-based "Save to notebook" on the Ask-AI popover, which only saves whatever passage you highlighted. Works from both the reading pane and the standalone report reader (/cafe/[cafeId]/artefact/[artefactId]). Owner-only — not shown on a shared/read-only report.

Every report carries an AI-generated content notice. A short line at the bottom of every report — "This report was generated with AI. AI can make mistakes — please verify important claims against the cited references." — is always shown, including on shared and read-only views, since the reminder matters most to a reader who isn't the researcher.

Every report type is now searchable and citable in Ask Your Cafe (W1.17). Previously only Auto-Research's extraction cards were embedded into the Q&A search index (because they were tray items, and tray items were embedded). Now every report type — Auto-Write, Detect Conflicts, Verify, Find the Truth, People research, and Auto-Research — is embedded and can ground a Grounded Q&A answer. Clicking a cited report reference in a Q&A answer opens that report in the reading pane, same as clicking a cited source. See Q&A Chat (Ask Your Cafe) below.

References are collapsible by default, on every report type (W55.1). The bibliography at the end of any research report — Auto-Write, Verify, Conflict, Find the Truth, or a person report — renders inside a collapsible disclosure, collapsed by default, so a long reference list doesn't dominate the page. Click it to expand. In-text [N] citation clicks still open (and expand) the right entry even while it's collapsed. Sources with no known publication date show "(accessed <date>)" instead of a bare "(n.d.)", and a source's URL is printed once rather than twice — both fixes apply everywhere citations render, on the web viewer and in the PDF. In the PDF specifically, the References section always starts on its own page, for every report type.

Tips

• Create multiple Cafes for different topics — one Cafe per research project
• Use Academic search for scholarly sources — try OpenAlex for broad coverage, arXiv for preprints, or PubMed for biomedical literature
• YouTube transcripts are great for extracting key points from lectures and interviews
• Combine sources — search the web, find papers, scrape articles, then curate the best bits on your Tray
• Analyze anything — the Analyze source button works on every saved source, so you can re-analyze content with different questions; or highlight any snippet and use the popover's Ask AI for a quick chat about just that part
• Multi-select presets — when analyzing YouTube or URLs, select multiple presets (Summarize + Key Points) to get several analyses at once
• Add Notes for external content — use the + button to paste in material from sources outside Coffeescribe (emails, documents, other tools)
• Combine related items — merge excerpts from different sources into a single consolidated reference before pasting into your Workspace
• Edit tray text freely — click any tray item to refine wording, trim irrelevant parts, or add your own annotations inline
• Ask your Cafe — once you have research, use the Ask Your Cafe chat (the reading pane's home view) to ask questions and get referenced answers grounded in your collected material
• Filter Q&A sources — click the funnel icon in the Q&A header to query only specific sources instead of your entire Cafe
• Save Q&A answers — save useful answers back to your Tray, as new sources, or to Notebooks for use in your Scribes. Citations are preserved in all three cases and remain clickable in-Cafe
• Rename sources — double-click any source title to give it a more descriptive name
• Share your research — use the Share button to send a read-only link to collaborators (Cafe sharing). They can browse and ask Q&A questions using their own tokens. To share a specific report only, use the Share button on the report card (see Auto-Write Research Reports)
• Paste a reading list in one go — got a tab-soup or a notes-app dump of links? Paste the whole blob into the search bar and the batch import modal opens automatically (up to 50 items per batch, every signed-in user). Way faster than feeding them in one at a time.
• Paste a DOI or PMID to get a fully enriched source card instantly — the identifier input auto-detects the type and fetches structured metadata so you don't have to copy-paste authors and journals manually (every signed-in user).
• Set your citation style once per Cafe — open Settings (gear icon), pick APA 7, MLA 9, Chicago N-B, or Harvard CTR, and every enriched source card renders formatted citations automatically.
• Check the credibility badge before citing — a High or Medium credibility source with "Academic Journal/Paper" tier and a green freshness badge is your safest bet for academic writing. Remember: the badge is a structural check, not fact-verification — always read the source.
• Enable Academic Sources Only in Settings — flipping this filter in a research Cafe for a dissertation or report hides web articles and social posts so only peer-reviewed and government sources appear.
• Try LinkedIn or Reddit Actors for professional or community insight — enable them via the Provider Picker; the one-time consent modal only appears on the first enable per Cafe (every signed-in user).
• Group related cafés on /research — drag one café onto another to create a group, or use a café's ⋮ menu → "Add to group…" on a phone. A café can belong to more than one group, grouping is free, and it never deletes anything — see Organizing Cafés into Groups.