Skip to content

Commit 8c3546d

Browse files
vdavidvdavid
committed
Search: Two-pass AI search with preflight
- LLM first generates a broad "preflight" query, runs it against the index, then refines using real results — fixes blind pattern guessing (e.g. `*resume*` matching `.fastresume`) - `translate_search_query` accepts optional `PreflightContext` (hit count + up to 50 sample entries with names/sizes/dates) for the refinement pass - `summarize_ai_query()` generates human-readable one-liners for the preflight row (e.g. `*resume* · size ≥ 10 MB · after 2025-01-01`) - `format_preflight_table()` and `build_refinement_system_prompt()` format results for the LLM's second pass - Preflight row always visible when AI is enabled — shows placeholder → summary → hit count → "Refining..." → "Refined" - Skips refinement when ≤10 hits (already precise enough) or 0 hits - Generation counter prevents stale LLM responses from overwriting newer queries - Pass 2 failure gracefully falls back to pass 1 results with "Refinement failed" indicator - UTF-8-safe name truncation in preflight table formatting - `PreflightContext`, `PreflightEntry` types, `preflightSummary` field on `TranslateResult` - 12 new Rust tests for summarization, table formatting, prompt construction, serde roundtrips
1 parent 0f2c25c commit 8c3546d

7 files changed

Lines changed: 684 additions & 49 deletions

File tree

0 commit comments

Comments
 (0)
⚡