RAG settings
RAG settings control how many knowledge-base chunks reach the model and how strict relevance matching is. They apply per assistant and take effect on the next visitor message.
Open RAG settings
Section titled “Open RAG settings”- Assistants → select assistant → Knowledge bases → RAG settings.
- Adjust values below.
- Save and re-test in the test panel.
Settings reference
Section titled “Settings reference”| Setting | Default | Range | Effect |
|---|---|---|---|
| similarity_threshold | 0.7 | 0.0 – 1.0 | Minimum vector similarity for a chunk to be included. Higher = stricter, fewer chunks. |
| max_results | 10 | 1 – 20 | Maximum chunks sent to the Response Synthesizer. |
| boost_recent | off | on / off | Weights newer documents higher when scores are close. |
| retention_days | plan default | varies | Drops indexed chunks older than N days from retrieval (does not delete source files). |
Query preprocessing and rewrite are handled by smart query understanding on Starter+ plans — not duplicated here.
Tuning guide
Section titled “Tuning guide”Too much irrelevant context
- Raise
similarity_thresholdto0.75–0.85 - Lower
max_resultsto5–8
Answers miss obvious content
- Lower threshold to
0.6–0.65 - Raise
max_resultstoward15–20 - Improve source documents — see Content best practices
Time-sensitive content (pricing, release notes)
- Enable boost_recent so the latest uploads rank higher
Interaction with hybrid search
Section titled “Interaction with hybrid search”Chatevo merges dense (semantic) and sparse (keyword) results before applying your threshold. Exact product SKUs and error codes still match via keyword even at a higher threshold.
See Document retrieval architecture.
Per-KB vs per-assistant
Section titled “Per-KB vs per-assistant”These settings are per assistant. Knowledge-base-level tuning (shared across assistants using the same KB) lives under Tune search settings.