Repointing Retrieval at ChromaDB

Where we left off

Lesson 20 introduced chromadb's API in isolation, one script, one query, side by side with what this course's own retrieve() would have returned. This lesson does the real swap: ask(), unchanged in shape since Lesson 8, now runs entirely on a chromadb backend instead of a Python list.

The interface stays the same, the implementation doesn't

def ask(query: str, collection: chromadb.Collection, k: int = 2) -> str:
    retrieved = retrieve(query, collection, k)
    return generate_answer(query, retrieved)

Compare this to Lesson 8's ask(). The signature is different (store: list[dict] became collection: chromadb.Collection), but the shape, retrieve then generate, is identical. This is exactly the point of building the hand-rolled version first: everything downstream of retrieval (prompting, generation, this course's whole mental model of what RAG is) never had to change when the storage backend did.

The code, piece by piece

def retrieve(query: str, collection: chromadb.Collection, k: int) -> list[str]:
    query_vector = embed_texts([query])[0]
    results = collection.query(query_embeddings=[query_vector], n_results=k)
    documents = results["documents"]
    assert documents is not None
    return documents[0]

Embedding the query is unchanged, still this course's own embed_texts. What's different is everything after that: instead of scoring every record by hand and sorting, collection.query() does it, and returns the matching documents directly (results["documents"][0], the [0] because chromadb supports querying with multiple query vectors at once, this lesson always passes exactly one).