[lszskye]

Purpose

Linked issue: #452
Implement precise file-index evaluation for mixed scalar and full-text predicates, with optional vector or full-text TopK applied after filtering.

• Add EnablePreciseFileIndexEvaluate for append-only reads and primary-key reads with deletion vectors.
• Evaluate AND/OR predicate trees and use available selectivity estimates to order AND children.
• Recheck scalar index candidates against the data.
• Apply predicates, deletion vectors and row ranges before root TopK, then push the final row selection into the data reader.
• Support multiple named score columns and keep scores aligned with selected rows.
• Reuse the index container through an instance-based FileIndexEvaluator.

Precise evaluation requires bitmap32-compatible file row positions and rejects bitmap64 deletion vectors and merge reads. Full-text and vector searches require suitable indexes; only scalar predicates support evaluation without an index. TopK remains per file.

Tests

read_context_test.cpp
empty_file_index_reader_test.cpp
file_index_reader_test.cpp
complete_index_score_file_batch_reader_test.cpp
file_index_evaluator_test.cpp
file_index_read_executor_test.cpp
internal_read_context_test.cpp
raw_file_split_read_test.cpp
table_read_test.cpp
predicate_abi_inte_test.cpp
read_inte_with_index_test.cpp
write_and_read_inte_test.cpp

API and Format

• Add ReadContextBuilder::EnablePreciseFileIndexEvaluate, defaulting to false, and carry the setting through ReadContext.
• Require precise evaluation and file-index.read.enabled=true for file-level vector/full-text searches and FullTextMatch predicates.
• Require explicit _<field>_INDEX_SCORE projections in precise mode; reject _INDEX_SCORE.
• Extend FileIndexReader with search capability checks and selectivity estimates. Full-text TopK returns scored results, and FullTextMatch supports optional scores.
• Add FileIndexer::CreateEmptyReader to preserve search capabilities for empty indexes.
• Handle FullTextMatch through FileIndexReader rather than the generic scalar FunctionVisitor.

Documentation

Update the read API comments and docs/source/api/file_index.rst with the precise evaluation option, search requirements and named score projection examples.

Generative AI tooling

Generated-by: OpenAI Codex (GPT-6)

[lxy-9602]

Please link an issue for this pr.

[SteNicholas]

@lszskye, I left some comments for this pull request. PTAL.

[SteNicholas]

@lszskye, thanks for updates. I left some comments for updates: an empty IN list on a Bloom-indexed column can crash, and a sub-millisecond bound can make precise evaluation drop a matching row.

[SteNicholas]

LGTM.

[lxy-9602]

+1