#5: Fuzzy/cross-field search + exact total hits

[davidkelley]

Goal

Match ES behavior for typo-tolerant multi_match and cross-field relevance, and allow callers to request exact total hit counts.

Features

• MultiMatch fuzziness (AUTO, numeric edit distance) for string queries.
• Combined/cross-fields scoring akin to ES combined_fields (BM25F style).
• Exact total hits toggle track_total_hits=true.

Implementation Plan

1) MultiMatch Fuzziness

• Extend QueryNode::MultiMatch to accept fuzziness (enum: Auto, Edits(u8)).
• Planner: when fuzziness set, expand terms using bounded Levenshtein per field’s analyzer tokens; cap expansions (max_expansions) and min_length similar to FuzzyOptions.
• Scoring: treat fuzzy expansions as additional terms with reduced boost (1 / (1+edit_distance)) or configurable.
• For single-term queries, allow Auto mapping: length 1–2 → 0, 3–5 → 1, 6+ → 2.
• Tests: typo cases (“pikchu” -> “pikachu”), ensure non-fuzzy default unaffected.

2) Combined Fields (Cross-Fields)

• Add match_type: CrossFields in MultiMatch already exists; implement BM25F-like logic:
  • Normalize term freq across listed fields; compute combined score using average field length.
  • Planner builds a term group spanning fields; scorer uses aggregated tf/len.
• Default operator: AND vs OR controlled by operator; support minimum_should_match.
• Tests: query spans name/set fields; verify relevance parity with ES behavior.

3) Exact Total Hits

• SearchRequest gains track_total_hits: Option<bool> (default false to preserve speed).
• Reader:
  • If flag true, compute exact doc count matching query (post-filter) without early termination; may reuse aggregation pipeline to count or run full collect with DocCollector.
  • If false, keep existing estimate.
• Response: reuse total_hits_estimate; when exact, set field and maybe add boolean total_hits_exact=true to signal accuracy (optional but recommended).

Code Touchpoints

• searchlite-core/src/api/types.rs: new fields/enums; serde defaults.
• searchlite-core/src/query/planner.rs & scorer: implement cross-field tf aggregation and fuzzy expansions.
• searchlite-core/src/api/reader.rs: track_total_hits handling; ensure cursors unaffected.
• HTTP layer: accept new JSON fields; validation errors on invalid fuzziness string.

Performance/Bounds

• Set per-request caps: fuzzy expansions max 50 terms per original token; fail fast if exceeded with 400.
• track_total_hits may be expensive; add warning log when enabled without limit.

Tests

• Unit: planner builds expected term groups; fuzzy expansion obeys limits.
• Integration: search request with track_total_hits=true returns exact count on small fixture.
• Regression: ensure profile/explain still work with new scoring paths.

Migration Notes

• Existing clients unaffected unless they opt-in to fuzziness/track_total_hits.
• Document default fuzzy behavior per Auto rules; provide examples mapping from ES queries used in managemco (multi_match with AUTO on single-term searches).