All Submissions:

⚠️ PRs must target the dev branch. If your PR targets master, please change the base branch to dev before requesting a review.

• My PR targets the dev branch (not master) and my branch was created from dev.
• Have you followed the guidelines in our Contributing document?
• Have you checked to ensure there aren't other open Pull Requests for the same update/change?

Changes to Core Features:

• Have you added an explanation of what your changes do and why you'd like us to include them?
• Have you written new tests for your core changes, as applicable?
• Have you successfully ran tests with your changes locally?

What this changes

This fixes a batch response-shape bug in the all-zero-limit fast paths for search, query, recommend, and discover.

For a single request, limit = 0 correctly means "return zero points", so the result is []. For a batch request, limit = 0 should only make that individual result slot empty; it should not remove the slot from the outer batch response.

Before this change, the all-zero-limit batch shortcuts returned vec![], so a non-empty batch could return zero outer result slots. This PR adds a small shared helper, empty_batch_results(request_count), and uses it in those shortcuts so each input request keeps a corresponding empty result vector.

Affected paths

REST/gRPC batch search/query/recommend/discover
  -> collection-level batch function
  -> every request has limit = 0
  -> previous fast path returned vec![]
  -> fixed fast path returns one empty result vector per input request

Changed collection-level fast paths:

Collection::core_search_batch(...)
Collection::do_query_batch(...)
recommend_batch_by(...)
discover_batch(...)

Only the all-zero-limit batch shortcut changes. Single-request zero-limit behavior still returns [], and mixed batches still use the existing non-shortcut path for shard routing, vector resolution, scoring, payload loading, and filtering.

Evidence

The batch cardinality invariant should be:

len(response.results) == len(request.requests)

For a two-request batch where both requests have limit = 0, the changed shortcut now preserves the outer result shape:

request: [limit=0, limit=0]

before: []
after:  [[], []]

The added focused regression test covers the shared helper shape used by the four fast paths:

empty_batch_results::<usize>(2) == [[], []]
empty_batch_results::<usize>(0) == []

This is source-level before/after evidence for the changed fast path. Local runtime validation is currently limited to git diff --check because the Rust toolchain is unavailable in the current Windows PATH and WSL environment.

Validation

• git diff --check
• Duplicate PR check: gh search prs --repo qdrant/qdrant --state open 'zero limit batch' and gh search prs --repo qdrant/qdrant --state open 'empty batch results' returned only this PR.
• Remote checks observed: CodeRabbit passed.
• cargo test -p collection empty_batch_results_preserves_outer_shape - not run locally because cargo is not available in the current Windows PATH or WSL environment.
• cargo +nightly fmt --all -- --check - not run locally because the Rust toolchain is unavailable.
• cargo clippy --workspace --all-targets --all-features -- -D warnings - not run locally because the Rust toolchain is unavailable.
• cargo build --workspace --tests --locked - not run locally because the Rust toolchain is unavailable.

AI assistance disclosure

Initial prompt intent for AI-assisted implementation and drafting:

Fix Qdrant's zero-limit batch result-shape bug so non-empty all-zero-limit batches preserve one empty result slot per input, then prepare a concise PR draft following Qdrant contribution rules.

Commit authorship disclosure:

- [AI-assisted] fix(collection): preserve empty result shape for zero-limit batches