Retrieval Tuning: Top-K, Filters, Re-Ranking, and Thresholds
Retrieval Tuning: Top-K, Filters, Re-Ranking, and Thresholds
Top-K selection
Too low = miss context. Too high = noise. Start with k=3–5 and tune using real queries.
Thresholds
If similarity is below threshold, don’t inject junk. Let the agent ask for clarification or use another source.
Re-ranking
Use a re-ranker to reorder candidates for better relevance, especially in large corpora.
