perf: use sets for O(1) membership testing in local search query path

[dubin555]

Description

The local search context-building code uses Python lists for entity/relationship/text-unit membership testing (in operator) across 10 call sites in 5 files. Since in on a list is O(n), this creates O(n·m) complexity in several hot paths during query execution. At scale (1000+ entities, 10000+ relationships), this adds over 1.7 seconds of unnecessary latency to the local search query path.

This PR converts list comprehensions to set comprehensions for all membership-test collections and builds defaultdict index dicts for two O(n·m) inner loops, reducing the combined hot-path complexity from O(n²) to O(n).

Related Issues

Related: #2250 (performance regression)

Proposed Changes

• 10 list→set conversions: Replace [...] list comprehensions with {...} set comprehensions wherever the resulting collection is only used for in membership testing:

  • local_context.py — selected_entity_names, entity name sets for relationship/covariate filtering
  • relationships.py — get_in_network_relationships, get_out_network_relationships, get_candidate_relationships, get_entities_from_relationships
  • covariates.py — get_candidate_covariates
  • community_reports.py — get_candidate_communities
  • text_units.py — get_candidate_text_units

• 2 algorithmic improvements using defaultdict(list) index dicts:

  • _filter_relationships() in local_context.py: relationship link counting now uses source/target index dicts instead of scanning all relationships per entity
  • build_covariates_context() in local_context.py: covariate filtering now uses a subject_id index dict instead of scanning all covariates per entity

Benchmark results

Standalone benchmark with isolated before/after implementations, all correctness assertions pass (results are identical between old and new code at every scale):

Scale	Function	List (ms)	Set (ms)	Speedup
100 entities, 1K rels	_filter_relationships	8.1	0.26	30x
100 entities, 500 covs	build_covariates_context	3.1	0.07	45x
500 entities, 5K rels	_filter_relationships	236	1.7	141x
500 entities, 2K covs	build_covariates_context	46	0.32	144x
1K entities, 10K rels	_filter_relationships	1048	3.6	291x
1K entities, 5K covs	build_covariates_context	246	0.86	286x

At 1000 entities / 10000 relationships, the combined hot path goes from ~1.7s to ~8.4ms.

Checklist

• I have tested these changes locally.
• I have reviewed the code changes.
• I have updated the documentation (if necessary).
• I have added appropriate unit tests (if applicable).

Additional Notes

All changes are mechanical — the in operator works identically on sets and lists, so there is no behavioral change. The two index-dict refactors produce identical results, validated by key-by-key comparison assertions in the benchmark at three different scales. No new dependencies are introduced.

[dubin555]

@microsoft-github-policy-service agree