Skip to main content
The most common retrieval question is: “I know the answer is in this document. Why can’t the Agent find it?” The retrieval page is the diagnostic tool. Run the query yourself, inspect the scores, and lock in fixes with an eval set so they don’t regress.

Dissecting a single query

Type the failing query on the retrieval page and hit Run.
  • Look at the left result list. Where does the chunk you expected rank, and what similarity score does it carry?
  • Pick a result and the right pane shows the full chunk, neighbouring chunks, and metadata. From here, Find similar, Open source file, Go to document, and Open refinery jump you straight to the likely cause.
  • Switch the method between Vector only and Keyword only and rerun. Which one misses tells you whether the query and the document mean different things, or share no literal terms.

Common failures and fixes

Eval sets

Whenever you fix a “why can’t I find X” issue, add it to an eval set so it doesn’t come back. Under Compare → Eval set:
  1. Add an item and fill in the query, expected substring, category (optional), and note (optional). Import in bulk with CSV.
  2. Run all to run every item against the current preset. Each row lights up pass/fail based on whether the expected substring appears in the results.
  3. Filter by category or status to zero in on the failures.
Compare mode eval set with options to add an item, import from CSV, or run all items
Add one item per fix. When you change presets, run the eval set in Compare mode first to catch regressions.