Where we left off

Lesson 12 filtered this course's own Python list by a source field before ranking by similarity. This lesson does the identical thing on a chromadb backend, using its built-in where clause instead of a list comprehension.

The code, piece by piece

metadatas=[{"source": path.name, "area": area} for path, area in zip(paths, areas)]

A second metadata field, area, alongside the source field from Lesson 20-21. Nothing about chromadb requires exactly one metadata field, any dict of extra values works, this lesson just needs a second one to filter by.

filtered = collection.query(
    query_embeddings=[query_vector],
    n_results=1,
    where={"area": "study"},
)

where={"area": "study"} narrows the candidate pool to only records whose area metadata equals "study", before ranking by similarity, same ordering of operations as Lesson 12's Python-level filter. This is the chromadb-native way of expressing the exact same idea, a query scoped to a subset of the collection instead of the whole thing.

Checkpoint

  • where clause: chromadb's built-in way to narrow a query by metadata before ranking, the same idea as Lesson 12's Python-level filter, expressed natively.
  • A real vector database's metadata filtering is typically far faster than filtering a Python list by hand at scale, since it can use its own indexes on metadata fields too, not just on the vectors.
  • Filtering to the wrong scope is still a real failure mode here, same as Lesson 12, a database doesn't fix a badly-scoped query, it just executes it faster.

If anything here still feels unclear, ask before moving to Lesson 23.