Skip to content

fix(indexer): align embedding batch size with encoder contract - #2

Merged
p3t3r67x0 merged 1 commit into
mainfrom
fix/encoder-batch-size
Oct 1, 2026
Merged

p3t3r67x0 merged 1 commit into
mainfrom
fix/encoder-batch-size

Conversation

@p3t3r67x0

Copy link
Copy Markdown
Collaborator

Knowledge could send up to 64 texts to /embed, while the existing Admin encoder accepts at most two, causing larger index reconcile runs to fail with 422 invalid_request.

Split missing embeddings into requests of at most two texts and enforce the same 1–2 text limit in Clients.embed(). Vector reuse, reconciliation, Qdrant upsert batches and delete ordering remain unchanged. The encoder repository is unchanged.

Regression tests cover multiple encoder requests for seven new chunks, correct vector/payload association, failure on the second embedding request before any upsert, and client batch-size bounds.

Validation:

  • pytest: 77 passed.
  • ruff check .: passed.
  • ruff format --check .: passed.
  • git diff --check: passed.

Follow-up to merged PR #1.

@p3t3r67x0 p3t3r67x0 self-assigned this Oct 1, 2026
@p3t3r67x0
p3t3r67x0 merged commit dbcc837 into main Oct 1, 2026
2 checks passed
@p3t3r67x0
p3t3r67x0 deleted the fix/encoder-batch-size branch October 1, 2026 21:29
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant