Ask your documents
Ask fact, comparison, and summary questions across your document collections and get an answer whose every claim quotes the passage it came from. When the evidence is weak, Colrows says so instead of guessing. Complete queries go further and read every document in a collection to count, sum, or list matches. To connect or upload documents first, see Document sources.
Where you ask
Document questions are asked in the document workspace, over the collections you select (up to 20 at a time). Each question runs under your own identity, so you only ever see answers drawn from documents you are allowed to read. Answering document questions inside the AI Data Analyst chat and the MCP tools is on the roadmap; today they are a separate surface.
Question types
| Type | Use it for |
|---|---|
| Automatic | Let Colrows pick the right mode for the question. |
| Fact | A single answer backed by specific passages - "what is the notice period in the Acme contract?" |
| Comparison | Compare a point across subjects, with evidence gathered from each in turn. |
| Summary | Summarise the selected collections. |
How an answer is built
- Colrows retrieves candidate passages with hybrid search - dense and sparse vectors - and reranks them.
- A model drafts the answer as separate claims, each carrying an exact quote and the element it came from.
- Every quote is checked in code to be a verbatim substring of the source. Claims whose quote is not found are dropped.
- A second, independent model pass reviews each remaining claim and removes any it cannot support.
- The platform verifies the answer once more before returning it.
Each claim shows a Source button that opens the cited passage: the quote, its page, heading path, and element, any table with its headers, and a link to download that exact version of the original document.
Answer statuses
| Status | Meaning |
|---|---|
| Answered | Supported claims, each with a citation. |
| Partial | Answered as far as the evidence goes, with the gaps and limitations listed. |
| Abstained | No answer: the evidence was missing or too weak, or a comparison had too few subjects. Colrows shows the closest passages so you can judge for yourself. |
| Unavailable | The question asks for something this mode cannot do safely - for example a count or "list all", which belong to a complete query (below). |
Complete queries: count, sum, and list
A ranked search answers "what does the evidence say?" It cannot answer "how many?" - the top results are not the whole set. Complete queries close that gap by reading every document in the selected collections, up to 1,000.
- Complete count - how many documents match.
- Complete sum - a total across the matching documents.
- Complete match list - the documents that match, with evidence.
Each document is given a definite outcome - Match, No match, Ambiguous, or Unreadable - with a quote for the decision. A final number is returned only once every document has a definite outcome, so a total is never built from a partial read. Progress and per-document outcomes are shown while the query runs.
This is the document equivalent of the SQL compiler's explicit error: an answer quotes its source, or Colrows tells you the evidence is not there. Answers are not combined with SQL query results today.