Metadata Filtering by Modality

Where we left off

Lesson 20 stored modality as chromadb metadata but never used it in a query. naive_rag Lesson 22 narrowed a chromadb query with where filters on source/area; this lesson applies the identical mechanism to modality, letting a caller ask for text-only results, image-only results, or both (the default), a scoping tool Lesson 13's Python-level balancing didn't have: "only images, ever, for this query" rather than "some of each, ranked separately."

Why this is different from Lesson 13's k-balancing

Lesson 13 guarantees representation from both modalities in one result set. This lesson does the opposite kind of thing: excluding a modality entirely, useful when a caller already knows the answer must be visual (a UI button labeled "search diagrams only", say) rather than letting the ranking decide. Both tools solve different problems; neither replaces the other.

The code, piece by piece

def query_collection(
    collection: chromadb.Collection,
    query_vector: list[float],
    n_results: int,
    modality: str | None = None,
) -> dict:
    where = {"modality": modality} if modality else None
    return collection.query(query_embeddings=[query_vector], n_results=n_results, where=where)

naive_rag Lesson 22's exact pattern: build a where clause only when a filter is actually requested (None means "no filter, search everything"), pass it straight to .query(). modality being one of exactly two string values ("text", "image") makes this filter trivial to reason about, unlike a filter over a large, open-ended set of source values.