Architecture Review: Run symbol searches on key interfaces to understand the dependency graph.
Plan Mode: Use this skill to populate the "Context" section of a Plan Mode artifact.
Refactoring: Identify all usages of a symbol before renaming or modifying it. </usage_patterns>

symbols "UserAuthentication"

Semantic Search:

search "authentication middleware logic"

</code_example>

RAG Evaluation

Overview

Systematic evaluation of RAG quality using retrieval and end-to-end metrics. Based on Claude Cookbooks patterns.

Evaluation Metrics

Retrieval Metrics (from .claude/tools/repo-rag/metrics.py):

Precision: Proportion of retrieved chunks that are actually relevant
- Formula: Precision = True Positives / Total Retrieved
- High precision (0.8-1.0): System retrieves mostly relevant items
Recall: Completeness of retrieval - how many relevant items were found
- Formula: Recall = True Positives / Total Correct
- High recall (0.8-1.0): System finds most relevant items
F1 Score: Harmonic mean of precision and recall
- Formula: F1 = 2 × (Precision × Recall) / (Precision + Recall)
- Balanced measure when both precision and recall matter
MRR (Mean Reciprocal Rank): Measures ranking quality
- Formula: MRR = 1 / rank of first correct item
- High MRR (0.8-1.0): Correct items ranked first

End-to-End Metrics (from .claude/tools/repo-rag/evaluation.py):

Accuracy (LLM-as-Judge): Overall correctness using Claude evaluation
- Compares generated answer to correct answer
- Focuses on substance and meaning, not exact wording
- Checks for completeness and absence of contradictions

Evaluation Process

Create Evaluation Dataset:

{
  "query": "How is user authentication implemented?",
  "correct_chunks": ["src/auth/middleware.ts", "src/auth/types.ts"],
  "correct_answer": "User authentication uses JWT tokens...",
  "category": "authentication"
}

Run Retrieval Evaluation:

# Using Python directly
from .claude.tools.repo_rag.metrics import evaluate_retrieval
metrics = evaluate_retrieval(retrieved_chunks, correct_chunks)
print(f"Precision: {metrics['precision']}, Recall: {metrics['recall']}, F1: {metrics['f1']}, MRR: {metrics['mrr']}")

Run End-to-End Evaluation:

# Using Python directly
from .claude.tools.repo_rag.evaluation import evaluate_end_to_end
result = evaluate_end_to_end(query, generated_answer, correct_answer)
print(f"Correct: {result['is_correct']}, Explanation: {result['explanation']}")

Expected Performance

Based on Claude Cookbooks results:

Basic RAG: Precision 0.43, Recall 0.66, F1 0.52, MRR 0.74, Accuracy 71%
With Re-ranking: Precision 0.44, Recall 0.69, F1 0.54, MRR 0.87, Accuracy 81%

Best Practices

Separate Evaluation: Evaluate retrieval and end-to-end separately
Create Comprehensive Datasets: Cover common and edge cases
Evaluate Regularly: Run evaluations after codebase changes
Track Metrics Over Time: Monitor improvements
Use Both Metrics: Precision/Recall for retrieval, Accuracy for end-to-end

References

RAG Patterns Guide - Implementation patterns
Retrieval Metrics - Metric calculations
End-to-End Evaluation - LLM-as-judge
Evaluation Guide - Comprehensive evaluation guide

Memory Protocol (MANDATORY)

Before starting: Read .claude/context/memory/learnings.md

After completing:

New pattern -> .claude/context/memory/learnings.md
Issue found -> .claude/context/memory/issues.md
Decision made -> .claude/context/memory/decisions.md

ASSUME INTERRUPTION: If it's not in memory, it didn't happen.