Evaluate: Measure Your RAG Pipeline's Quality

Scope: This skill focuses on answer quality metrics (faithfulness, relevancy, correctness) against a labeled test set. For operational performance metrics (latency, throughput, cost), use /benchmark.

Run a structured evaluation of the learner's RAG pipeline using established metrics. This skill teaches evaluation methodology while generating actionable results.

Step 1: Identify the Pipeline

Welcome! Let's see how your RAG pipeline is performing.

First, check whether the learner has existing work to evaluate:

• Look for a learner profile at progress/learner-profile.md and for code in src/ and projects/.
• If pipeline code exists in projects/, great — proceed. If multiple pipelines exist, ask which one to evaluate.

• If no pipeline or RAG code exists anywhere in projects/ or src/, guide them warmly:

  "It looks like you haven't built a pipeline yet — that's totally fine! Let's get you set up first. Run /build to create your first RAG pipeline, and then come back here to see how it scores. It only takes a few minutes to get something running!"

  Stop here — do not continue to Step 2.

• If a pipeline is found, verify it has the minimum components: a retriever and a generator.

Step 2: Explain the Evaluation Framework

Before running metrics, teach the learner what they are measuring and why:

Core RAG Metrics

• Faithfulness: Does the generated answer stick to the retrieved context? (measures hallucination)
• Answer Relevancy: Is the answer actually relevant to the question asked?
• Context Precision: Are the retrieved documents relevant to the question?
• Context Recall: Did the retriever find all the relevant information?

Additional Metrics (if applicable)

• Answer Correctness: How close is the answer to a ground truth answer?
• Latency: How long does each pipeline stage take?
• Token Usage: How many tokens are consumed per query?

Explain each metric with a simple analogy so the learner builds intuition.

Step 3: Prepare Test Data

Help the learner create or use evaluation data:

1. Generate test queries: Create 10-20 representative questions for their corpus.
2. Create ground truth (if needed): For each question, write the ideal answer and list the relevant source documents.
3. Save the test set to projects/[pipeline-name]/eval/test-set.json.

If they already have a test set, load and validate it.

Step 4: Run the Evaluation

Walk through running the evaluation step by step:

1. Set up the evaluation framework (RAGAS, custom, or both).
2. Run each test query through the pipeline, capturing: the query, retrieved contexts, generated answer, and latency.
3. Compute metrics for each query.
4. Aggregate results.

Provide code snippets the learner can run. Explain what each part does.

Step 5: Present Results

Display results in a clear, readable format:

RAG Pipeline Evaluation Report
==============================
Pipeline: [name]
Test queries: [N]
Date: [today]

Overall Scores:
  Faithfulness:      [score] / 1.0
  Answer Relevancy:  [score] / 1.0
  Context Precision:  [score] / 1.0
  Context Recall:     [score] / 1.0

Avg Latency: [X]ms (retrieval: [Y]ms, generation: [Z]ms)

Highlight the weakest metric and explain what it means practically.

Step 6: Improvement Recommendations

Based on the results, provide specific, actionable suggestions:

• Low faithfulness -> check prompt template, add grounding instructions, consider smaller chunks
• Low context precision -> revisit chunking strategy, try re-ranking, tune embedding model
• Low context recall -> check chunk coverage, try hybrid search, increase top-k
• High latency -> profile each stage, consider caching, batch embeddings

Prioritize the top 2-3 improvements that would have the biggest impact.

Step 7: Save and Track

Save the evaluation report to projects/[pipeline-name]/eval/reports/report-[date].md. Update progress/module-tracker.md to note the evaluation was completed. Encourage the learner to re-evaluate after making improvements to see their progress.

Suggest 2-3 relevant next steps using slash commands:

• /benchmark — measure operational performance (latency, throughput, cost) alongside quality
• /debug-rag — diagnose and fix the weakest areas identified in your evaluation
• /build — implement the improvements suggested above and re-evaluate