pgvector: HNSW or IVFFlat index for 50k embeddings?

HNSW, for your size. Reasons:

• IVFFlat builds its clusters from the data present at index time. Built on an early or empty table, the clusters are poor and recall drops, which is the "missing matches" you see. You'd have to rebuild it as data changes.
• HNSW has better recall at the same speed and doesn't need rebuilding.

create index on chunks using hnsw (embedding vector_cosine_ops);

Match the operator class to the operator you query with (<=> is cosine). Building it on 50k rows needs some maintenance_work_mem; on a small compute it can take a few minutes, run it outside peak hours.

HNSW here too. If you stay on IVFFlat for some reason, raise ivfflat.probes (default 1) at query time, e.g. set local ivfflat.probes = 10. With probes = 1 you search only one cluster, which explains the missing matches.

Small one: OpenAI embeddings are already normalized to length 1, so inner product (<#> with vector_ip_ops) gives the same ranking as cosine and is a bit cheaper. Not worth changing if cosine is fast enough.