Ask — Cross-paper RAG¶
The Ask workspace at /ask lets you pose questions that are answered from your entire library at once, rather than from a single paper. It uses a retrieval-augmented generation (RAG) pipeline that searches across all your embedded paper chunks, re-ranks results, and streams a grounded answer with citations.
The chat interface¶
Composing a question¶
Type your question in the input box at the bottom of the workspace and press Enter or click Ask. There is no limit on question length, but concise, focused questions produce better results.
Questions can be:
- Factual — "What datasets does the literature use to evaluate X?"
- Comparative — "How do methods A and B differ in their treatment of Y?"
- Exploratory — "What are the open problems in Z according to the papers in my library?"
Streaming answers¶
Answers stream token-by-token via SSE (server-sent events). A loading indicator is shown while the system is retrieving and ranking chunks; the answer text then appears progressively as the model generates it.
Each answer message contains:
ConfidenceBadge — an indicator of how well the retrieved source passages support the answer, measured by sentence-level matching. There are four levels:
| Badge | Meaning |
|---|---|
| Verified | Every checkable sentence in the answer matched a source passage |
| Mostly verified | At least half of the checkable sentences matched |
| Partially verified | Some checkable sentences matched |
| Unverified | None of the checkable sentences matched |
Unverified ≠ wrong. It means the system could not match the answer's sentences to the retrieved source passages — the answer may still be accurate, but it could not be verified against what was retrieved from your library.
“Verified” is also not independent fact-checking: it describes textual support in the retrieved passages, not whether the paper or generated interpretation is scientifically correct.
Click the badge to see which sentences could not be matched to a source.
SourcesAccordion — collapsed by default; click to expand. Lists the Source Passages retrieved and used to construct the answer. Each entry shows the paper title, the passage excerpt, and a link to the Paper Detail page for that paper.
FeedbackButtons — thumbs-up and thumbs-down buttons for rating the answer quality. Your ratings are recorded and can inform future improvements to the retrieval pipeline.
Follow-up questions carry the current conversation context, so you can build on a previous answer without repeating background information.
Message history¶
Conversation messages persist across navigation within the same session using the chat store. If you navigate away to another page and return to /ask, your previous messages are still visible. The history is cleared when you sign out or start a new session.
Differences from single-paper Ask¶
The Ask workspace at /ask searches across all embedded papers in your library. The in-paper Ask section on the Paper Detail page is scoped to a single paper's chunks. Use the cross-paper Ask workspace when you want an answer synthesised from multiple sources.
Offline behaviour¶
The Ask workspace requires a live connection to the LLM backend. It is not available offline. If you are offline, the input is disabled and a notice explains that cross-paper RAG is unavailable.
Related pages¶
- Paper Detail — single-paper Ask tab for questions scoped to one document.
- Research Feed & Library — add more papers to your library to broaden the evidence base for answers.
- Settings — configure the LLM models used for answering.