[lll-hhh]

EmbeddingClient.get_embedding([a, b, c]) returns vectors in response-array order, ignoring each embedding's index. An OpenAI-compatible endpoint returning indices [2, 0, 1] therefore associates every vector with the wrong input.

Sort batch response entries by their input index before extracting vectors. Single-string requests and the existing failure fallback retain their behavior. The regression uses a real OpenAI SDK CreateEmbeddingResponse with shuffled entries; the ordered-response fixture now also supplies the SDK's index field.

Validation on Windows / Python 3.12:

• The shuffled-response regression fails on unchanged main (0fa8c4d); the eight existing embedding tests pass there.
• OPENAI_API_KEY=test-key-for-unit-tests python -m unittest tests.test_embedding tests.test_embedding_api_base tests.test_novelty_rejection tests.test_novelty_response_parsing -v: 23 passed.
• All applicable pre-commit hooks (isort, black) and git diff --check pass.
• Ran python -m unittest discover tests on both head and untouched base in the same environment. Head ran 1,075 tests and base ran 1,074. Both reported 24 failures, 94 errors and 8 skips; the 118 failing test/subtest identities are identical after normalizing generated temporary-directory names. The added regression passes on head. The complete suite is not green locally; no new failing tests were introduced in that comparison.

No real embedding requests or model inference were used. AI assistance: Codex investigated, implemented and validated this change.