fix: sort unscored retriever results last

[linhongyu510]

What this PR does

Fixes #13089.

MultiQueryEmbeddingRetriever, MultiQueryTextRetriever, and TextEmbeddingRetriever used doc.score or 0.0 when sorting results. That places score=None alongside zero and therefore above valid negative retrieval scores.

This PR follows Haystack's existing convention in _deduplicate_documents, DocumentJoiner, and AnswerJoiner: a missing score sorts as negative infinity, while every numeric score—including 0.0 and negative values—keeps its real ordering.

The change covers both sync and async execution paths for all three retriever wrappers. It does not modify the underlying retrievers or their scoring.

Why this matters

Negative retrieval scores are supported. In particular, unscaled BM25Okapi may return them. Before this fix, combining or wrapping such results could promote an unscored document above a genuinely scored match:

before: positive, unscored, zero, negative
after:  positive, zero, negative, unscored

Tests

Added paired sync/async regression tests for:

• MultiQueryEmbeddingRetriever
• MultiQueryTextRetriever
• TextEmbeddingRetriever

Each test mixes positive, zero, negative, and missing scores.

Verification performed:

• UV_EXCLUDE_NEWER=false hatch run test:unit <six affected test files>
  • 60 passed, 12 deselected
• Six focused regression tests
  • all 6 fail on the pre-fix implementation
  • all 6 pass with this change
• Targeted mypy check
  • no issues in 9 source files
• Ruff check
  • all checks passed
• git diff --check
  • clean

UV_EXCLUDE_NEWER=false was only needed locally because the repository's 24-hour supply-chain cutoff filtered the already-published build dependency while resolving against the current system date; no dependency or project configuration was changed.

Scope

• No public API change.
• No score normalization or score mutation.
• Numeric ordering remains unchanged.
• Only None moves after all numeric scores.

AI assistance

Developed with AI-assisted tooling. I reviewed and understood the final changes and independently ran the reproducer, counterfactual regression test, unit tests, lint, and type checks.

[vercel]

@linhongyu510 is attempting to deploy a commit to the deepset Team on Vercel.

A member of the Team first needs to authorize it.

[github-actions]

Coverage report

Click to see where and how coverage changed

File	Statements	Missing	Coverage	Coverage (new stmts)	Lines missing
haystack/components/retrievers
multi_query_embedding_retriever.py
multi_query_text_retriever.py
text_embedding_retriever.py
Project Total

This report was generated by python-coverage-comment-action