Documentation Index
Fetch the complete documentation index at: https://koreai.mintlify.app/llms.txt
Use this file to discover all available pages before exploring further.
This guide covers validating your Search AI configuration through testing and debugging tools to ensure optimal answer quality before deployment.
Navigation: Answer Generation > Test Answers
Testing Answers
Access Testing
- Navigate to Configuration > Answer Generation.
- Click Test Answers
- Enter a query
- Review the generated answer
- Use debug option to analyze behavior
Testing Workflow
| Step | Action | Purpose |
|---|
| 1 | Enter test query | Simulate user input |
| 2 | Review answer | Verify response quality and accuracy |
| 3 | Open debug view | Understand how answer was generated |
| 4 | Analyze chunks | Check which content was used |
| 5 | Refine configuration | Adjust settings based on findings |
Debug Components
| Component | Description |
|---|
| Qualified Chunks | Chunks selected and used to generate the answer |
| Retrieval Details | How chunks were identified and ranked |
| LLM Request/Response | Full prompt sent and response received (for generative answers) |
| Processing Time | Time taken by each component |
Agentic RAG Debugging
When Agentic RAG is enabled, an additional Retrieval tab appears showing:
| Information | Description |
|---|
| Agent Sequence | Order in which agents were invoked |
| Agent Input | Data sent to LLM by each agent |
| Agent Output | Results returned from each agent |
| LLM Timing | Time taken per LLM call |
Answer Insights
The Answer Insights feature provides analytics for query-response interactions.
Available Data
| Feature | Description |
|---|
| Query Grouping | View all answers for grouped queries |
| Search Logs | Filter logs by answer and channel |
| Detailed View | Query overview, debug info, LLM details |
| Performance Tracking | Monitor answer quality over time |
Accessing Answer Insights
Navigate to Analytics > Search AI > Answer Insights. Learn More.
Debugging Checklist
Common Issues and Solutions
| Issue | Possible Cause | Solution |
|---|
| No results returned | Content not indexed | Verify content sources and extraction settings |
| Poor relevance | Threshold too high/low | Adjust similarity score threshold |
| Missing information | Chunks too small | Increase chunk size or token budgets |
| Incomplete answers | Insufficient context | Increase Top K chunks or token budget |
| Business rules not applying | Condition mismatch | Test with debug to verify rule triggers |
| Slow responses | Too many LLM calls | Review Agentic RAG agent usage |
Configuration Verification
| Check | Location | What to Verify |
|---|
| Retrieval Strategy | Configuration > Retrieval | Vector vs. Hybrid selection |
| Thresholds | Configuration > Retrieval | Similarity, proximity, Top K values |
| Answer Type | Configuration > Answer Generation | Extractive vs. Generative |
| LLM Settings | Configuration > Answer Generation | Model, prompt, temperature |
| Business Rules | Configuration > Business Rules | Active rules and conditions |
Best Practices
Testing Strategy
- Test incrementally - Validate each configuration change before moving to the next
- Use varied queries - Test different query types, lengths, and phrasings
- Include edge cases - Test ambiguous queries and boundary conditions
- Compare results - Document before/after when making changes
Debug Analysis
- Review qualified chunks - Ensure relevant content is being selected
- Check chunk rankings - Verify highest-ranked chunks are most relevant
- Analyze LLM prompts - Confirm context is properly structured
- Monitor timing - Identify performance bottlenecks
Ongoing Monitoring
- Track Answer Insights - Review analytics regularly
- Monitor feedback - Enable user feedback and review ratings
- Iterate configuration - Continuously refine based on data
- Document changes - Keep records of configuration modifications
Testing Scenarios
Scenario 1: Basic Answer Validation
1. Enter simple factual query
2. Verify answer accuracy
3. Check source citation
4. Confirm response time acceptable
Scenario 2: Retrieval Quality Check
1. Enter query matching specific content
2. Open debug view
3. Verify expected chunks are qualified
4. Check similarity scores
Scenario 3: Business Rule Verification
1. Configure test rule with known conditions
2. Enter query that should trigger rule
3. Open debug view
4. Confirm rule was applied correctly
Scenario 4: Agentic RAG Testing
1. Enable Agentic RAG
2. Enter complex query
3. Review Retrieval tab in debug
4. Verify agent sequence and outputs
Quick Reference
Debug Tab Contents
| Tab | Shows |
|---|
| Qualified Chunks | Selected content for answer |
| Retrieval | Agent processing (Agentic RAG only) |
| LLM Details | Prompt and response data |
Key Metrics to Monitor
| Metric | Healthy Range |
|---|
| Response Time | < 3 seconds (varies by LLM) |
| Chunk Relevance | Top chunks match query intent |
| Answer Accuracy | Matches source content |
| User Feedback | Positive ratings trending up |
