Tune KB hit quality (Confidence Threshold)
You have indexed your documentation, the chunk counter shows e.g. 580 indexed chunks, but when asked „How does the chat work?" the bot responds with a generic „How can I help you?" and the Copilot displays:
No hit in the knowledge base — this suggestion is generic.
This is almost always due to a single value: the Confidence threshold for KB hits.
What is Confidence?
During retrieval, the AI compares the visitor's question (as a vector embedding) with each indexed KB chunk. Each chunk receives a Similarity Score between 0.0 (no relevance) and 1.0 (text-identical).
Only chunks above the Confidence threshold are considered hits and included in the context that the AI uses to generate the answer. Everything below is discarded — even if it would actually fit.
The Default Value: 0.75
The standard is 0.75. This is strict. Consequence:
- Word-for-word formulations match safely
- Synonyms and paraphrases (typical visitor questions!) often fail
- Short questions („How to chat?") have weak embeddings and almost always fail
For most knowledge bases with natural language articles, 0.75 is too high.
Adjust the Slider
- Settings → AI Chatbot → „Knowledge" tab
- When „Use knowledge database" is enabled, the Hit Threshold slider appears (0.00 to 1.00)
- Move the value, Save
Changes take effect immediately — no reindex required, only the retrieval query changes.
Recommended Values
| Value | Character | When Useful |
|---|---|---|
| 0.3 | Very loose | Very small KB, every hit counts; many false positives acceptable |
| 0.5–0.6 | Loose | Recommended for most setups — catches paraphrases and short questions |
| 0.75 | Strict (Default) | Large KB with clear article titles, precision more important than recall |
| 0.85+ | Very strict | Only text-similar matches; almost like a FAQ search |
Rule of thumb: Rather too loose than too strict. A false hit is usually filtered out by the AI prompt („Answer only if the information is in the context"), a missing hit makes the bot silent.
Interaction with „Handoff at Low Confidence"
Under Settings → AI Chatbot → „Handoff" tab there is the toggle „Handoff at Low Confidence". This takes effect as soon as no chunk reaches the threshold.
This means:
- Low threshold (0.4) + Handoff active → Bot tries RAG answers for many questions; Handoff only if really nothing fits
- High threshold (0.9) + Handoff active → Bot frequently hands over to an agent, even for questions that the KB could actually answer
- Low threshold + Handoff off → Bot almost always answers with something from the KB; possibly inaccurate
The combination determines the perceived character of your bot („helpful" vs. „quickly hands over to humans").
If Nothing is Found Even at 0.4
If the slider is set to 0.4 and the bot still finds nothing, it's probably due to:
- Widget KB Scope — the widget has mode „Selected Categories" with categories in which the articles are not located → no hit possible. Check the live counter in the Widget „Knowledge" tab
- Language Mismatch — articles are in English, the question in German (or vice versa). The embedding vector is language-sensitive, cross-language retrieval works only weakly
- Articles not published — only
status = publishedcounts for RAG. Drafts are ignored - Index outdated — new articles created but never indexed. Settings → AI Chatbot → „Knowledge" tab → „Reindex"
Per-Widget Override
Currently, the threshold is tenant-wide. Per-widget overrides are prepared as a feature stub in the save path, but not exposed in the UI — if you need different strictness per widget, contact us.
Tracking
Every time the threshold is not reached, the bot logs the question as a KB Gap (if gap tracking is active). You can see these under Knowledge Base → Gaps — ideal for knowing which articles are missing in your KB.