Define the multimodal search retrieval contract
Start with the questions multimodal search must answer, the evidence each answer must cite, and the region where that evidence belongs. These constraints determine index boundaries and key scopes before data is written.
Write a small labeled query set before tuning. It prevents a faster or cheaper configuration from silently reducing evidence quality.
- Name the user decision each result supports
- Store stable source and version metadata
- Separate read credentials from ingestion credentials
Apply the search latency guide to the data path
For multimodal search, make ingestion idempotent and keep a durable job record for every source. Dense and lexical representations serve different query language, so measure each path and the combined result.
Filters should express real authorization and business boundaries. They are not an after-the-fact relevance patch.
- Use regional indexes
- Keep metadata filterable
- Evaluate dense, lexical, and hybrid retrieval
Operate with evidence
Track request count, error rate, p50/p90/p99, and relevance on the same dashboard. When a result changes, compare the source version and query set before changing ranking.
Publish the limits of the corpus to users. A retrieval system is trustworthy when it can say what it did not find.