Semantic document search

Search by meaning, not by filename

Semantic search finds the document that answers your question even when nobody used your words for it. Ask whether a large customer can walk away if the company is acquired, and the room returns the agreement whose clause is headed "Assignment", plus everything else covering the same idea.

• No filename guessing. MSA_final_v3_JB.pdf and Copy of Schedule 4.docx are as findable as anything else.
• No contract vocabulary required. A keyword search for "change of control" misses the agreement that says "Change in Control of the Company", then points at "Section 14.3" for forty pages.
• Strongest for the side that did not build the room. Buy-side teams and their advisers read someone else's filing system on a deadline.
• Useful on the sell side too. Answer a diligence request by finding the documents that bear on it, without walking the tree.

Results stop at your permissions

Search has no access rules of its own. It resolves your access through the same per-folder permission code path as the file tree, and the check runs before a document is admitted, not as a filter over a ranked list.

• A document in a folder you cannot read never appears in results and never affects the ranking.
• Each deal is its own database on its own subdomain, so a query on one deal cannot reach another deal's files.
• Read access and export rights stay separate. A view-only document can be returned and opened in the viewer while download, native-file streaming and bulk ZIP stay closed, checked per file inside a ZIP.

More on the security page.

How it works

The room is searched by meaning rather than by wording. Every indexed document gets a title, a summary and a text transcription, and that text becomes an embedding: a numeric fingerprint of what it is about. Your query gets one too, and results rank by closeness of meaning.

• Uploads never wait on AI. The file is stored and visible right away. Indexing runs as a detached background job, so a slow or failing model cannot stall an upload or lose a file.
• Text is pulled locally first. Office formats are read through their internal XML, and CSV and plain text go in directly. Images go to a vision-capable model. A PDF is sent to the model as a file when the local pass returns little, which is reading, not OCR.
• Ranked, with a floor. You get the closest matches above a similarity threshold. If nothing clears it, nothing comes back. No padding with weak hits.
• Failure is soft. If indexing fails, the document still exists, still opens and still carries a filename-derived title.

Questions

Does the AI chat use the same index? Yes. Chat retrieval and the scoring of new uploads against open diligence requests run on the same vectors, behind the same permission check. That is why the room can propose which request a new upload answers.

Search returned nothing. Now what? Ask chat the same question. Chat applies a looser similarity floor, so it sometimes answers from a document search did not return.

A file is in the room but never appears in results. It probably never finished indexing, which leaves it viewable but invisible to search, chat and request matching. Nothing on the document says so, which is a real gap. The advisory-side deal team can run a reprocess action that re-indexes every document lacking a vector. Run it after a bulk load.

Where it stops

• Scans. There is no OCR step. A PDF that is a photograph of paper, with no text layer, is indexed on little more than its filename. Load searchable PDFs where you can.
• Spreadsheets. Cell text, sheet names and headers extract well. Formulas and cross-tab structure do not reduce to prose, so a model is found through its labels.
• Long files. One vector per document, embedded text truncated, no passage-level chunking. A clause buried in a long agreement may not pull it to the top, and a hit points at the document, not the paragraph.
• Exact terms. There is no keyword fallback. If you know the filename or a verbatim phrase, exact matching is the better tool.
• Very large rooms. Every query compares against every document vector, with no approximate index. Sound at normal deal-room volumes, not for tens of thousands of files.
• View-only is not unquotable. Anything you may read, chat may quote back to you. If a document is too sensitive to quote, close the folder.

Semantic search builds a candidate set fast. It does not prove that set is complete, so where completeness is the point, use it to start and check against the folder structure. More on the features page.