Skip to main content

Exercise 2 - Full Pipeline

Goal: Wire retrieval and generation into a single ask(question) function. You now have a complete, working naive RAG system.

Background​

You have all the pieces:

  • FAISS index with embedded chunks (Foundations 03)
  • A retriever that returns top-k chunks (Naive RAG 01)
  • A prompt builder that grounds the LLM (this module, Exercise 1)

This exercise combines them into one callable pipeline.

Assignment​

Open 02_full_pipeline.py.

  1. Load your FAISS index and chunks.
  2. Write ask(question, k=5) that:
    • Retrieves top-k chunks for the question
    • Builds the strict prompt from Exercise 1
    • Calls llama3.2 for the answer
    • Returns the answer string
  3. Run it against at least 5 questions - mix of in-corpus and out-of-corpus.
  4. For each answer, also print which chunks were retrieved (the first 80 chars of each). Verify the answer is supported by what was retrieved.

Stretch: Add a verbose=True parameter to ask that prints the full prompt before sending it to the model. Read the prompt carefully - does it look the way you expect?

Thinking questions​

  • You built the index once (in Foundations) and reuse it here. In a real application, when would you need to rebuild it?
  • Your pipeline makes two network calls per question: one to the embedding endpoint, one to the generation endpoint. Which is slower? Does it matter?
  • If the retriever returns a wrong chunk (low quality retrieval), the generator will produce a wrong answer confidently. What would you add to this pipeline to catch that?

← Exercise 1 · Next: 03 Cloud Models →