Advanced search options allow you to optimize how the assistant finds and uses information in your knowledge base. These parameters directly influence response quality and system performance.
How to access search options
- Select the assistant you want to configure
- In the sidebar, click on "Knowledge Base"
- Scroll to the "File Search Options" section
- Modify parameters according to your needs
- Save changes to apply new settings
Available parameters
Maximum number of results
- What it controls: How many content fragments are retrieved from the knowledge base for each question
- Typical range: From 5 to 20 results
- Default value: 10 results
- Impact: More results = more complete answers but higher token consumption
Relevance score threshold
- What it controls: How relevant content must be to be considered useful
- Range: From 0.00 (includes everything) to 1.00 (only extremely relevant content)
- Default value: 0.70
- Impact: High threshold = fewer but more pertinent results
How these parameters work together
The system uses both parameters in sequence:
- Initial search: Finds all potentially relevant content
- Relevance filter: Discards content below the score threshold
- Numerical limitation: Takes only the first N most relevant results
- Final processing: AI uses this content to formulate the response
Configuration examples
"Precise" configuration (FAQ/Support)
- Number of results: 5-8
- Relevance threshold: 0.80-0.90
- When to use: Precise answers to specific questions
- Advantages: Focused answers, lower token consumption
"Complete" configuration (Consulting)
- Number of results: 15-20
- Relevance threshold: 0.60-0.70
- When to use: Complex questions requiring information from multiple sources
- Advantages: More complete and detailed answers
"Balanced" configuration (General use)
- Number of results: 10-12
- Relevance threshold: 0.70-0.75
- When to use: Most use cases
- Advantages: Good compromise between quality and performance
Factors to consider in configuration
Knowledge base size
- Small base (few files): Lower threshold to find sufficient results
- Large base (many files): Higher threshold to filter noise
Content type
- Specialized content: Higher threshold for precision
- General content: Lower threshold for coverage
Expected question types
- Specific questions: Fewer results, high threshold
- Exploratory questions: More results, moderate threshold
Effectiveness monitoring
- Regular testing: Try typical questions with different configurations
- User feedback: Collect opinions on response quality
- Conversation analysis: Verify if assistant always finds relevant information
- Iterative optimization: Gradually modify parameters based on results