RAG settings¶
Retrieval tuning
RAG settings control how much context the agent receives and how relevant that context must be. Too strict returns nothing; too broad adds noise.
Tune retrieval with real questions from callers, not abstract guesses.
Configure how your agent retrieves and uses knowledge base content.
Settings Reference¶
| Setting | Description | Default | Range |
|---|---|---|---|
ragTopK |
Number of documents to retrieve | 5 | 1-20 |
ragSimilarityThreshold |
Minimum similarity score (0-1) | 0.70 | 0.0-1.0 |
Top K (ragTopK)¶
Controls how many document chunks are retrieved.
| Value | Behavior |
|---|---|
| 1-3 | Very focused, highest relevance only |
| 5-10 (recommended) | Good balance of relevance and coverage |
| 10-20 | Broad coverage, may include less relevant |
Example
With ragTopK: 5, the system retrieves the 5 most similar document chunks.
Similarity Threshold (ragSimilarityThreshold)¶
Filters out less relevant matches.
| Value | Behavior |
|---|---|
| 0.9+ | Only very close matches (may return nothing) |
| 0.7-0.9 (recommended) | Good relevance filter |
| 0.5-0.7 | More permissive, includes broader matches |
| 0.5- | Very permissive, may include irrelevant content |
Example
With ragSimilarityThreshold: 0.70, only documents with similarity ≥ 70% are included.
Finding the Right Balance¶
Too Strict (High threshold, low K)¶
- Symptom: Agent doesn't use KB information
- Fix: Lower threshold, increase K
Too Permissive (Low threshold, high K)¶
- Symptom: Agent gets irrelevant or conflicting information
- Fix: Raise threshold, decrease K
Recommended Starting Point¶
Adjust based on your specific use case and document quality.
Query Transformation¶
For better retrieval, enable query transformation (see Agent Options):
{
"queryTransformEnabled": true,
"queryTransformHistoryLimit": 5,
"queryTransformConfidenceThreshold": 0.8
}
This analyzes conversation context to reformulate queries for better matching.