Retrieval and Quality Tuning

Hit Testing (retrieval lab) compares retrieval results without directly modifying the production configuration. It shows stage information for recall, fusion, rerank, and final ranking.

Running a Query

1. Open the knowledge base details and select Hit Testing.
2. Enter a query of 1-1000 characters.
3. Select the retrieval method: vector, full-text, or hybrid.
4. Set Top K (1-20, default 5), dense threshold (0-1, default 0), weights, and rrf_k (1-1000, default 60).
5. Override rerank and the candidate pool size (1-100) as needed.

A/B Comparison

A/B comparison accepts up to 10 configurations, and each configuration ID must be unique. Results preserve the success or failure status of each side; a failure on one side does not hide the other side's successful results.

Reading Diagnostics

Diagnostics may indicate that the knowledge base is not enabled, an embedding model is missing, a timeout occurred, or processing failed. Stage timings are split into recall, rerank, context, and total. Hybrid scores use weighted reciprocal rank fusion: Σ weight × 1/(rrf_k + rank).

Applying Presets

After saving a local preset, select Apply to Production. Before applying, confirm the embedding model, index dimensions, and weights; at least one hybrid retrieval weight must be greater than 0.

Common failures and recovery steps are described in the Knowledge Base settings reference.