Bundata Docs
Relevance Recall (finding the right docs) Precision (avoiding wrong docs) Troubleshooting poor recall or precision RAG and agents Next steps

Vector Search best practices

Use these practices to get better relevance, recall, and precision from Vector Search and to troubleshoot when results are poor.

Relevance

  • Right collection — Query the collection that contains the document set for your use case (contracts, invoices, policies). Avoid querying a single “everything” collection without filters.
  • Metadata filtering — Use filters (source, date, document type) to scope results. See Filtering.
  • Chunk quality — Smart bites should be meaningful units (e.g. a section, a clause). Poor chunking (e.g. mid-sentence splits) hurts relevance. Tune extraction and chunking in your pipeline. See Extraction overview and Partitioning & chunking.
  • Freshness — Keep the catalog updated with workflow orchestration so new and updated documents are searchable. Stale data leads to missing or outdated context. See Vector Catalog: Lineage.

Recall (finding the right docs)

  • Increase k — Retrieve more candidates (e.g. top 10 or 20) and then re-rank or filter in your app if the answer isn’t in the top 3.
  • Broaden filters — If you over-filter (e.g. narrow date range), relax filters to see if the right document exists in the collection.
  • Check indexing — Ensure the document was successfully extracted and ingested. Use source lineage and run history to confirm. See Extraction runs.

Precision (avoiding wrong docs)

  • Tighten filters — Restrict by document type, source, or date so irrelevant docs are excluded.
  • Confidence — Filter or down-rank low-extraction confidence bites if they cause wrong context. See Extraction best practices.
  • Surface lineage — Show source lineage in the UI so users can ignore or report bad sources. See Agents grounding.

Troubleshooting poor recall or precision

  1. Confirm the document is in the collection — Check run history and lineage; verify the collection and schema.
  2. Inspect chunks — Look at the smart bites for that document. If chunking is wrong (e.g. too small or too large), adjust extraction/chunking settings.
  3. Try the query — Run the same query in the Platform or API and inspect returned chunks. Add or relax filters and re-run.
  4. Check embedding model — Ensure query and index use the same model; confirm the model fits your language and domain.

RAG and agents

  • Pass enough context — Send enough retrieved chunks to the LLM so the answer can be grounded in the right passage. Balance token limits and relevance.
  • Cite sources — Always attach source lineage to the response so users can verify grounded answers. See Grounding from catalog.

Next steps

Previous
Filtering