Relevance Threshold

The relevance threshold is a filter that determines how pertinent content must be to the user's question to be considered useful. It is expressed as a decimal value from 0.00 to 1.00, where higher values indicate greater selectivity.

How the relevance score works

The system assigns a score to each content fragment based on:

• Semantic similarity: How conceptually similar the content is to the question
• Lexical correspondence: Presence of specific keywords
• Thematic context: Belonging to the same knowledge domain
• Content structure: Relevance of titles, subtitles and structure

Interpretation of threshold values

Low threshold (0.30-0.60)

Characteristics:

• Includes even vaguely related content
• Greater information coverage
• Risk of non-pertinent information

When to use:

• Limited knowledge base
• Highly specialized content
• Exploratory and research questions
• Initial configuration for testing

Practical example:

Medium threshold (0.60-0.80)

Characteristics:

• Balance between coverage and precision
• Filters marginally relevant content
• Good general response quality

When to use:

• Standard configuration for most cases
• Well-organized knowledge base
• General assistant use

Practical example:

High threshold (0.80-1.00)

Characteristics:

• Only highly pertinent content
• Very focused responses
• Risk of missing useful information

When to use:

• FAQ with precise answers
• Very large knowledge base
• Content with possible overlaps
• When precision is critical

Practical example:

Optimization strategies

Gradual approach

1. Start with medium threshold (0.70): Safe value for most cases
2. Monitor responses: Check if they are too generic or too limited
3. Adjust incrementally: Change by 0.05-0.10 at a time
4. Test with real questions: Use actual user questions

A/B testing for optimization

• Prepare test question set: 20-30 representative questions
• Compare different thresholds: 0.60, 0.70, 0.80
• Evaluate each response: Quality, completeness, relevance
• Choose winning threshold: The one with best balance

Common problems and solutions

Problem: "Assistant never finds information"

Possible cause: Threshold too high

Solution:

• Reduce threshold by 0.10-0.15
• Verify content quality in knowledge base
• Check that content uses consistent terminology

Problem: "Responses include too much non-pertinent information"

Possible cause: Threshold too low

Solution:

• Increase threshold by 0.10-0.15
• Remove duplicate or very similar content
• Improve content organization

Problem: "Inconsistent responses"

Possible cause: Contradictory content in relevance range

Solution:

• Increase threshold to exclude marginal content
• Identify and resolve content contradictions
• Better organize knowledge base

Optimization for specific sectors

Technical support

• Recommended threshold: 0.75-0.85
• Objective: Precise and unambiguous procedures
• Considerations: Avoid contradictory information that confuses users

E-commerce

• Recommended threshold: 0.65-0.75
• Objective: Balance product information and related suggestions
• Considerations: Include similar products and accessories

Professional services

• Recommended threshold: 0.60-0.70
• Objective: Complete coverage of complex topics
• Considerations: Allow connections between related concepts

Continuous monitoring

• Quality metrics: Collect feedback on response relevance
• Conversation analysis: Identify patterns of unanswered questions
• Periodic testing: Repeat optimization every 2-3 months
• Content adaptations: Recalibrate when adding new materials