Search

When you submit a query, the system runs a multi-stage pipeline:

1. Query reformulation — optionally rewrites your query using HyDE, Multi-Query, or Step-Back strategies to improve recall.
2. Hybrid retrieval — combines vector similarity (semantic meaning) with full-text keyword matching. Typically retrieves 20+ candidates.
3. Permission filtering — removes documents the user doesn't have access to in the source system.
4. Document expansion — when multiple top results come from the same document, pulls in the full document for complete context.
5. Reranking — a second-pass scorer (BM25, Cohere, or cross-encoder) re-scores candidates and selects the top 5.
6. AI generation — the LLM reads all top sources and writes an answer, citing only the sources it finds useful.
7. Entity linking — if the Knowledge Graph is enabled, matching entities appear alongside the answer as clickable cards.

Use the source filter dropdown to restrict your search to specific data sources. This is useful when you know which system contains the information.

Click More Filters next to the source filter to refine your search by:

• File Type — filter by document format (PDF, Word, Excel, HTML, etc.). The list shows only file types present in your index.
• Date Modified — restrict results to documents modified within a specific date range.

Facet filters are applied at the retrieval stage, so the AI only considers matching documents when generating answers.

Each answer includes numbered citations. Click a citation to view the original document, page, and source location.

When the Knowledge Graph is enabled, search results may include entity cards — compact cards showing people, organizations, technologies, or other entities mentioned in your query. Each card shows the entity type, mention count, and related documents. Click a card to explore its relationships in the Knowledge Graph.

After each search result, the system suggests related follow-up questions. Click a suggestion to run it as a new search. This feature can be toggled in Settings → AI Assistant → Search Pipeline.

The percentage on each source represents its relevance score — how closely it matched your query based on keyword overlap and semantic similarity. The top-scoring result is always 100%, and others are relative.

Click the pin icon on any history item to keep it at the top. Pinned searches survive history clearing and are always accessible for one-click re-runs.