Course Wiki
RAG and Agent Evaluation — LLM Zoomcamp Module 4
--- video_url: 'https://www.youtube.com/watch?v=VKHBP0QSCFo&list=PL3MmuxUbc_hLZFNgSad56pDBKK8KO0XIv' --- # RAG and Agent Evaluation
Related Pages
LLM Zoomcamp › Module 4: Evaluation › RAG and Agent Evaluation
Notes
video_url: “https://www.youtube.com/watch?v=VKHBP0QSCFo&list=PL3MmuxUbc_hLZFNgSad56pDBKK8KO0XIv”
RAG and Agent Evaluation
So far, we evaluated retrieval. We checked whether search returns the document that should answer the question.
That is only the first step. A complete application still needs to produce a final answer. For RAG, this means checking the generated answer. For agents, it also means looking at the tool calls the model made before producing the answer.
RAG evaluation checks the whole flow together.
This includes:
- search
- prompt
- LLM
If the final answer is bad, the problem can come from any of these steps. The search might retrieve the wrong document, the prompt might omit important context, or the LLM might ignore the context.
In this part, we’ll evaluate:
- RAG answers with an LLM judge
- Agent answers and tool-call trajectories
We won’t go deep into agent evaluation frameworks here. We’ll use the agent from module 01, save the final answer, and also save the tool calls. Then we can look at whether the answer is good and whether the trajectory looks reasonable.
LLM as a judge
For RAG and agent evaluation, we compare the generated answer with the original answer. The generated answer won’t use the same words as the original. It’s a generative model, so the phrasing will be different even when the meaning is the same.
This is why we use another LLM to do the comparison. We show the judge the question, the original answer, and the generated answer. Then we ask it to decide if they are semantically equivalent.
This approach is called LLM-as-a-judge. The evaluating LLM is the judge. It classifies each answer as good or bad and explains its reasoning. Asking the judge to explain why it made a decision generally produces better classifications than asking for just the verdict.
Next, we’ll start with the RAG case and generate answers for the ground truth questions.
Key concepts
Related notes
- llmz-m01-agents
- llmz-m01-building-the-prompt
- llmz-m01-data-ingestion
- llmz-m01-function-calling
- llmz-m01-introduction
- llmz-m01-other-frameworks
- llmz-m01-quick-rag-revision-optional
- llmz-m01-rag
- llmz-m01-rag-helper
- llmz-m01-search
- llmz-m01-the-agentic-loop
- llmz-m01-the-course-faq-dataset
- llmz-m01-the-llm
- llmz-m01-toyaikit
- llmz-m01-wrap-up-of-part-1
- llmz-m02-embeddings
- llmz-m02-next-steps
- llmz-m02-rag-with-vector-search
- llmz-m02-vector-search
- llmz-m02-vector-search-with-minsearch
- llmz-m02-vector-search-with-pgvector
- llmz-m02-vector-search-with-sqlitesearch
- llmz-m03-ai-agents
- llmz-m03-ai-copilot
- llmz-m03-ai-orchestration
- llmz-m03-best-practices
- llmz-m03-context-engineering
- llmz-m03-multi-agent-systems
- llmz-m03-next-steps
- llmz-m03-retrieval-augmented-generation
- llmz-m04-agent-evaluation
- llmz-m04-evaluation
- llmz-m04-generating-ground-truth-data
- llmz-m04-generating-ground-truth-for-all-documents
- llmz-m04-generating-rag-answers
- llmz-m04-llm-as-a-judge
- llmz-m04-next-steps
- llmz-m04-search-evaluation
- llmz-m04-search-evaluation-metrics
- llmz-m04-search-parameter-tuning
- llmz-m05-assistant
- llmz-m05-built-in-judge
- llmz-m05-capturing-metrics
- llmz-m05-chat-app
- llmz-m05-docker-compose
- llmz-m05-monitoring
- llmz-m05-next-steps
- llmz-m05-querying-data
- llmz-m05-storing-data-in-postgresql
- llmz-m05-streamlit-dashboard
- llmz-m05-synthetic-data-generation
- llmz-m05-user-feedback
- llmz-m06-best-practices-for-rag
- llmz-m06-document-reranking
- llmz-m06-next-steps
- llmz-m07-chunking-for-longer-texts
- llmz-m07-content-processing-cases-and-steps
- llmz-m07-end-to-end-project-example
- llmz-m07-evaluating-rag
- llmz-m07-evaluating-retrieval
- llmz-m07-interface-and-ingestion-pipeline
- llmz-m07-monitoring-and-containerization
- llmz-m07-summary-and-closing-remarks