fix(retriever): skip FAISS padding slots instead of returning the last document

[winter-street]

Related Issues

• fixes FAISS backend returns the last document for result slots it could not fill #517

Proposed Changes:

FAISS pads its result with -1 when the index holds fewer vectors than the
requested top_k. BaseIndexBackend.contents is a plain list
(index_backends/base.py:30), so self.contents[-1] wraps around to the last
document instead of signalling a miss: the padding came back as duplicated
passages, indistinguishable from real hits downstream.

The change skips those slots:

for doc_id in doc_ids:
    if doc_id == -1:
        continue
    cur_ret.append(self.contents[doc_id])

MilvusIndexBackend and QdrantIndexBackend are unaffected — they resolve
content from the payload the store returns rather than by positional index into
contents.

How did you test it?

Ran it, with faiss-cpu 1.15.1. A 3-passage index queried with top_k=5:

	FAISS raw indices	search() returns
before	[0, 2, 1, -1, -1]	['DOC_A', 'DOC_C', 'DOC_B', 'DOC_C', 'DOC_C']
after	[0, 2, 1, -1, -1]	['DOC_A', 'DOC_C', 'DOC_B']

This PR adds tests/servers/retriever/test_faiss_search_padding.py, four cases
covering the padding, top_k below the index size, top_k equal to it, and a
single-passage index:

uv sync --extra retriever
uv run pytest tests/servers/retriever/test_faiss_search_padding.py -v
# 4 passed

With the fix reverted, the two padding cases fail
(expected 3 real hits, got 5: ['DOC_A', 'DOC_C', 'DOC_B', 'DOC_C', 'DOC_C'])
and the two normal-path cases still pass, so the tests pin the behaviour rather
than the implementation.

uv run pytest tests/ servers/retriever/tests servers/generation/tests could
not be collected in my environment (openai, litellm not installed), so I
have not run the rest of the suite — the change is confined to one module.

Notes for the reviewer

PR #425 proposed the same guard, bundled with five other fixes across five
files, and was closed. I opened this as a single-file change with a regression
test in case the bundling was what sank it. If the team's reason for closing
that PR applies to this fix as well, please say so and I will close this one
rather than resubmit.

Checklist

• One concern per PR, no unrelated reformatting.
• Regression test added under tests/servers/retriever/.
• ruff check and ruff format --check clean on the new test file. (faiss_backend.py reports seven pre-existing rules — I001, UP035, UP045, UP006 — on lines this change does not touch; I left them alone rather than bundle unrelated edits.)

[xhd0728]

Thanks for the focused fix.