Search

The Search tab is a built-in playground for querying a knowledge base the same way a workflow or agent would. Use it to sanity-check that your documents were ingested well and that a given question retrieves the chunks you expect — before you wire the knowledge base into anything.

Running a query

1. Open a knowledge base and switch to the Search tab.
2. Type a natural-language query. Search is semantic: it matches on meaning, so "how do I cancel" will find a chunk titled "Terminating a subscription" even though no words overlap.
3. (Optional) Add a metadata filter as JSON to restrict which chunks can match.
4. Submit. Results come back ranked by relevance.

Reading the results

Each result card shows:

• The source document the chunk came from.
• The chunk index — which slice of the document matched.
• A relevance score from 0–100%, where higher means a closer semantic match.
• The chunk's metadata, including the file name and anything attached during ingestion.

If a query you expect to work returns nothing useful, check that the relevant document is Ready (see Add documents), and consider whether the chunk settings are splitting the content too finely or too coarsely.

Metadata filters

The metadata filter narrows the candidate set before ranking. It accepts a JSON object whose keys are matched against each chunk's stored metadata, for example:

{"file_name": "refund-policy.pdf"}

Only chunks whose metadata satisfies the filter are eligible to be returned. This is the playground equivalent of the metadata and tag filters you'll use from workflows and agents.