Skip to content

Tune search settings

Knowledge-base search settings control how Chatevo retrieves chunks during RAG. They apply at the KB level; assistants can add overrides in RAG settings.

Settings reference

Setting	What it does	Starting point
similarity_threshold	Minimum vector similarity for a chunk to be included	`0.7` — raise if answers pull weak matches; lower if answers miss obvious content
max_results	Maximum chunks returned to the Response Synthesizer	`5`–`8` for focused FAQs; `10`–`15` for long policy docs
boost_recent	Prefer newer snapshots when multiple versions exist	Enable when docs update frequently (release notes, pricing)
retention_days	How long old snapshots are kept before purge	Align with your compliance policy; longer retention preserves audit history

Step 1 — Open search settings

Go to Knowledge bases and open a knowledge base.
Open Search settings (or Retrieval).
Adjust values and Save.

Step 2 — Test changes

Open an assistant with this KB attached.
Use Test chat with questions that previously missed or over-fetched context.
Iterate on similarity_threshold and max_results until citations look right.

How settings interact with hybrid search

Retrieval always runs dense + keyword search in Qdrant, merged with RRF. Threshold and max results apply after fusion:

Query Strategist rewrite
    → Hybrid search (dense + keyword → RRF)
    → Filter by similarity_threshold
    → Trim to max_results
    → Optional boost_recent reorder
    → Chunks to Response Synthesizer

On Starter+ plans, the Query Strategist can rewrite vague questions before search. See Smart query understanding.

When to tune per KB

KB type	Suggestion
Short FAQ	Higher threshold, fewer max results
Large manuals	Slightly lower threshold, more max results
Frequently updated site crawl	Enable boost_recent; pair with Auto-retrain on Standard+