Exercise 3 - ChromaDB

Goal: Replicate the FAISS search from Exercise 1 using ChromaDB. Experience the difference: persistence and metadata filtering come for free.

---

Background

FAISS is explicit - you manage vectors, you manage persistence, you manage metadata. ChromaDB wraps all of that into a collection. The trade-off: less control, more convenience.

---

Assignment

Open 03_chromadb.py.

1. Create a chromadb.PersistentClient pointed at ./chroma_db.
2. Get or create a collection called "rag-exercises".
3. Load and split the same PDF from previous exercises.
4. Add all chunks to the collection. Each chunk needs:
   • A unique id (e.g. "chunk_0", "chunk_1", ...)
   • The document text
   • The embedding (you still generate this with Ollama)
   • metadata with at least {"source": filename, "chunk_index": i}
5. Query the collection with the same queries from Exercise 1. Print the top-3 results.

Bonus: Add a metadata filter to your query. For example, only return results where chunk_index > 10. Does the result set change?

---

Thinking questions

• In Exercise 2, you had to manually manage two files (index.faiss and chunks.json). What does ChromaDB give you in exchange for that bookkeeping?
• ChromaDB can also generate embeddings for you if you pass an embedding_function. Why does this curriculum generate embeddings explicitly instead?
• When would you choose FAISS over ChromaDB for a production system?

---

← Exercise 2 · Next: Naive RAG →