How CompanyOS Search Finds the Right Passage

August 13, 2026

CompanyOS retrieves. Your AI client answers.

A Member asks a question in Claude, ChatGPT, or another compatible AI client. The client may call 3 read-only tools: search, read, and list. It can search visible knowledge, read a file by the key search returned, then write its reply.

CompanyOS does not synthesize that reply. It retrieves relevant evidence from the files visible to the current Member, and source files stay read-only throughout.

Five lists inspect the same files

Search is not a single index. CompanyOS runs these five lists in parallel, each already limited to the caller’s Workspace and Personal Files. That is the same model as the docs.

List 01

Chunk full text

Lexical match on passage text.

List 02

Chunk embeddings

Nearest-neighbor match on stored passage vectors.

List 03

Title and path

Lexical match on titles, folders, and heading paths.

List 04

Summary full text

Lexical match on the Distilled Summary, when one exists.

List 05

Summary embeddings

Vector match on that abstract, when one exists.

A Distilled Summary is a retrieval aid. It never gates a file becoming readable, and it is never the cited artifact. Citations point at the Document Version.

Chunk full text

If someone asks for the “launch checklist,” a file that contains those words can surface even when its title is different.

Chunk embeddings

Each passage is stored as a vector. The question is embedded the same way, then nearest neighbors are ranked. This catches strong meaning matches that do not repeat the exact query words.

Title and path

Titles, folders, and heading paths get their own lexical list, so a Document named “Launch Plan” can still win when the question uses different words in the body.

Distilled Summaries

After a Document is Ready, CompanyOS can add a document-level Distilled Summary used only to improve retrieval recall. Citations always point at the Document Version, never at the summary itself.

Separate candidates become one ordered list

The five lists run in parallel. CompanyOS fuses them with Reciprocal Rank Fusion so a passage that only won on meaning can still compete with one that matched the exact words. A deterministic pass can then prefer recency and metadata matches. Last, a configured rerank model scores the fused candidates against the question and reorders what the AI client receives.

Retrieve in parallelFuse and rankRerank for the question

A retrieval failure does not become a permission failure

If embeddings or summaries are unavailable, search continues on the lists that still returned hits and marks the result as degraded. Results may be less complete, but the system can still return grounded passages.

Authorization does not degrade. If permission evidence is missing, ambiguous, or stale, CompanyOS fails closed. An out-of-scope file never appears as a partial hit.

What search does not do

  • CompanyOS does not train on Workspace files.
  • Search covers only knowledge the current Member can already see: Company Files plus that Member’s Personal Files.
  • AI clients cannot edit or delete source files through these tools.
Get startedAll posts