Document Intelligence
Answers from your documents. Quoted, governed, complete.
Permissions enforced, every answer quoted from the source, and a clear “not found” when the evidence isn’t there.
How it works
Connect
S3, Google Drive, or OneDrive, or upload files directly.
Parse
Parsed locally, with OCR for scans and structure kept for tables.
Respect permissions
Each document keeps its permissions, checked on every request.
Ask & verify
Every claim quotes its source and is verified — or Colrows abstains.
Four reasons it reads differently
Permission-aware by default
Answers only from documents a person could already open. Stale access fails closed.
Quoted, verified answers
Every sentence quotes its source. If the quote isn’t there, the sentence is dropped.
Complete answers, not top-ten guesses
Ask “how many contracts have a change-of-control clause?” — Colrows reads every document, not the ten closest passages.
Documents strengthen your semantic layer
Policies and specs propose governed definitions to approve, and flag where a document contradicts a metric.
Ask in the way the question needs
| Question type | Use it for |
|---|---|
| Fact | A single answer backed by specific passages. |
| Comparison | Compare a point across subjects, with evidence from each. |
| Summary | Summarise the selected collections. |
| Complete count / sum / match list | Read every document (up to 1,000) to count, total, or list matches. |
Every answer comes back as Answered, Partial, or Abstained.
Sources, formats & security
Sources & formats
S3, Google Drive, and OneDrive (read-only), plus upload. PDF, DOCX, Markdown, and text up to 32 MiB; Google files read as PDF.
Security
Read-only connectors, per-user permissions, fail-closed on stale access, tenant isolation, retention and legal hold, and answers that are never cached.
Learn more in the docs: Document sources, Ask your documents, and Collections & permissions.
What’s next. Document answers are asked in the document workspace today. The AI Data Analyst chat and MCP tools, combined SQL + document answers, and more connectors (SharePoint, Confluence) are on the roadmap.
Questions? We have answers.
How is this different from RAG?
Plain RAG retrieves passages and lets a model write the answer, which can be confidently wrong. Colrows holds retrieval to a higher standard: every claim in an answer quotes an exact passage that is checked in code, a second model reviews each claim, and the platform verifies the answer before returning it. When the evidence is weak, Colrows abstains instead of guessing.
Can it count across all my documents?
Yes. A ranked search cannot answer “how many?” because the top results are not the whole set. Complete queries read every document in a collection, up to 1,000, give each a definite outcome, and return a total only once every document has been decided.
Which sources and file types are supported?
Connect AWS S3, Google Drive, or Microsoft OneDrive as read-only sources, or upload files. Colrows reads PDF, DOCX, Markdown, and text up to 32 MiB per file; Google Docs, Sheets, and Slides are read as PDF. SharePoint, Confluence, and other sources are on the roadmap.
Who can see which documents?
A person only gets answers from documents they are allowed to read. Collections have explicit readers, connected documents are mapped to individual users, and group or public sharing never grants access. Stale permissions fail closed, and access is checked on every request.
Is this in the AI Data Analyst chat?
Document questions are asked in the document workspace today. Answering them inside the AI Data Analyst chat and the MCP tools, and combining document answers with SQL results, are on the roadmap.
Put a question to your documents.
See cited answers, complete queries, and per-user permissions on your own documents.