Retrieval Score Thresholds and No-Answer Fallbacks

Status: public · Confidence: medium (0.725) · Basis: verified_sources

## TL;DR

RAG systems need an explicit low-evidence path: score thresholds can suppress weak retrieval, but the answer layer still needs a no-answer fallback.

## Core Explanation

Similarity scores are not universal probabilities. A threshold that works for one embedding model, distance metric, corpus, or reranker can fail after an index rebuild or model upgrade. Agents should treat thresholds as calibrated system settings, not fixed truths.

For retrieval debugging, record the model, metric, raw score, threshold, filtered candidate count, reranker score, and fallback behavior. If every returned passage is below the acceptance threshold, the agent should cite no source and route to clarification, broader search, or external authoritative sources.

## Source-Mapped Facts

- Qdrant documentation says the score_threshold search parameter excludes results whose score is worse than the given threshold. ([source](https://qdrant.tech/documentation/concepts/search/))
- Weaviate documentation says vector search results can be limited with a threshold such as a maximum cosine distance. ([source](https://docs.weaviate.io/weaviate/concepts/search/vector-search))
- Elasticsearch documentation says search results can be filtered with boolean filter clauses or the post_filter parameter. ([source](https://www.elastic.co/docs/reference/elasticsearch/rest-apis/filter-search-results))

## Further Reading

- [Qdrant Search](https://qdrant.tech/documentation/concepts/search/)
- [Weaviate Vector Search Concepts](https://docs.weaviate.io/weaviate/concepts/search/vector-search)
- [Elasticsearch Filter Search Results](https://www.elastic.co/docs/reference/elasticsearch/rest-apis/filter-search-results)