Docs

Search

Find ranked sources, cited summaries, and relevant connections from the same search in CompanyOS and your AI client.

Last verified: 2026-09-29

Home and connected AI clients use the same CompanyOS search. Ask a question to get ranked sources, a cited summary, and connections supported by those sources. Search includes Company Files and your own Personal Files, with matching published procedures and templates listed separately.

Read a search result

Enter your question on Home and select Search. In a connected AI client, ask it to search CompanyOS and cite its sources.

  1. Read the summary for context. Each paragraph identifies supporting sources.
  2. Open a source to inspect the parsed document text. CompanyOS reads the document’s Markdown, not the original DOCX or PDF bytes.
  3. Inspect the connections to see how the returned evidence relates. An unrelated workspace overview never replaces an empty search graph.

Search returns up to ten document results by default, plus up to ten matching published procedures and templates. These published items have their own source numbers. A rank score shows reciprocal rank: it is not a confidence percentage.

How CompanyOS ranks documents

Search combines word matches and meaning across passages, titles, paths, and document summaries. Each retrieval path is restricted to your workspace and visible files before ranking.

CompanyOS combines the ranked lists, then uses the configured reranking model when available. If part of retrieval fails, it returns the available evidence with a degraded status. If retrieval is unavailable, it reports an error instead of claiming that no evidence exists.

The answer summary uses only returned evidence, matching published content, and supported graph relationships. If summarization is unavailable, you can still read the sources. Document summaries used during retrieval are distinct from this answer summary; citations identify the underlying document evidence.

The illustration below walks through one fictional question.

Use the search tools

Connected AI clients can use these read-only tools:

Tool Purpose
search Search with an optional cited summary, ranked sources, related published content, and relevant connections
fetch Read a returned documentId as canonical Markdown; continue with continuationOffset for long documents. A file that has not produced canonical text returns not-ready guidance
list_files Browse visible files and their processing states

Search does not fetch every live integration automatically. Your AI client can discover connected Apps with find_apps and call their operations with call_app.

In an AI client, CompanyOS returns a short search status and supplies the assistant with the sources, cited context, graph relationships, and suggested next steps. The assistant can explain the findings in the conversation; no embedded search screen is required. Connecting CompanyOS to a host is separate from publishing it in an app directory.

Understand visibility and recovery

Search never includes another member’s Personal Files. Graph connections retain source evidence and follow the same visibility rules. Valid graph suggestions publish during processing without approval; manual graph edits remain protected.

If a file is still preparing, inspect its processing status in Files. If a ready file has older pending graph suggestions, use Refresh graph in graph settings, or Redo graph for that file. Queued work is not proof that processing has finished.

For permissions and file states, read Files. For connection setup, read AI Clients.

Search includes visible relationships up to two connections away from the matching sources. The chart can show up to 120 nodes and 200 relationships. Private material, deleted files, and superseded document versions remain excluded. Connected context helps explain the matches; it does not turn the chart into an unrelated workspace overview.

Illustrative example. A fictional question with conceptual timing, not a live search request.

Typical question runs the full pipeline to cited results.

QuestionScope gatePassage wordsPostgreSQL full-textPassage vectorsmeaning searchTitle / pathmetadata matchSummary wordsPostgreSQL full-textSummary vectorsmeaning searchRank fusionDeterministic orderModel rerankCited sourcesGraph contextProceduresAnswer summaryResults

Step 1 of 10