Retrieval Configuration
Fine-tune how your agent retrieves and uses knowledge
Retrieval Configuration#
Control how your agent searches and uses its knowledge base when answering questions.
What is Retrieval?#
When a user asks a question:
- Search: The agent searches the knowledge base
- Retrieve: Relevant content is retrieved
- Rank: Content is ranked by relevance
- Generate: Agent uses top content to answer
Retrieval Configuration lets you tune each step.
Accessing Settings#
- Go to your agent dashboard
- Navigate to Settings → Retrieval Config
- Adjust settings as needed
- Save changes
Configuration Options#
Chunk Limits#
Control how much content is retrieved:
| Setting | Description | Default |
|---|---|---|
| Docs Max Chunks | Knowledge base content | 5 |
| Examples Max Chunks | Example responses | 3 |
| Style Max Chunks | Learned style content | 2 |
| Conversation Max Chunks | Past conversation context | 3 |
Priority Weights#
Control which content types are prioritized. Each content type (docs, examples, style, conversation) has an adjustable priority weight. Higher priority means content is more likely to be retrieved.
Similarity Threshold#
Minimum relevance score for retrieved content:
- Higher: Only very relevant content (fewer results, more precise)
- Lower: Include somewhat related content (more results, broader coverage)
Advanced Options#
Hybrid Search#
Combine semantic and keyword search:
- Vector Weight: How much weight to give semantic understanding
- BM25 Weight: How much weight to give exact keyword matching
Reranking#
Re-score results after initial search:
- Enable Reranking: Use secondary scoring
- Top K: How many to rerank
Multi-Query#
Generate multiple search queries for better coverage:
- Enable Multi-Query: Search with variations
- Query Count: Number of variations (2-5)
Presets#
Quick configurations for common scenarios:
| Preset | Description |
|---|---|
| Balanced | Default settings for most use cases |
| Precise | Higher threshold, fewer but more accurate results |
| Broad | Lower threshold, more comprehensive coverage |
| FAQ-focused | Prioritize Q&A examples |
When to Adjust#
Increase chunks if:
- Agent misses relevant information
- Questions span multiple topics
- Knowledge base is large
Decrease chunks if:
- Responses include irrelevant info
- Agent mixes up topics
- Responses are too long
Raise threshold if:
- Agent uses unrelated content
- Answers wander off-topic
Lower threshold if:
- Agent says "I don't know" too often
- Related content isn't being found
Best Practices#
- Start with defaults: Adjust only if needed
- Test changes: Verify with sample questions
- Document changes: Track what you modified
- Reset if unsure: Use presets to restore defaults