Response - Retrieval confidence threshold

Introduction

Retrieval confidence threshold is a global T3CS setting that controls how strong a vector match must be before T3AC (chatbot) treats retrieved content as reliable enough for RAG (retrieval-augmented generation).

It replaces a former hard-coded value (0.62) with a configurable option in AI Foundation → AI Features → Training.

Set Retrieval confidence threshold under AI Foundation → AI Features → Training (Embeddings pipeline — ns_t3ac).

Property reference

Property	Value
Setting key	retrievalConfidenceThreshold
Allowed range	0.40 – 0.90 (values outside range are clamped)
Default	0.62

Purpose

Reduce answers built on weak or unrelated chunks (hallucination risk), while avoiding false “no results” when the best match is good but reranking lowered the displayed score below cosine similarity.

Tuning guide

Threshold	Effect
0.58 – 0.60	More answers; good when top matches often sit around 0.60–0.65
0.62 (default)	Balanced strictness
0.65 – 0.70	Stricter; fewer weak-context answers, more “no confident match” behavior

Note

Changing the threshold does not require re-training embeddings; flush caches and retest queries.

How the score is calculated

For each retrieved chunk, the gate uses:

Item confidence = max(score, similarity)

• similarity — cosine similarity between query and chunk embedding

• score — hybrid / reranked value (can be lower than similarity)

The top confidence among accessible chunks is compared to retrievalConfidenceThreshold.

Confidence gate passes if either

• Top confidence ≥ threshold, or

• Term-anchored bypass: at least one accessible chunk has confidence ≥ 0.45, confidence ≥ 0.38 (absolute floor), and at least one extracted query term appears in chunk content or path (word boundaries; path/hyphen-aware).