* Graph-derived prompt matching for the front-load hook's MEDIUM tier: * which indexed symbols do these prose words name? "state machine des * commandes" → `OrderStateMachine`, in any human language whose technical * nouns are Latin script — no keyword list involved. * * Precision com
(words: string[], limit: number = 6)
| 2781 | * returned symbol is guaranteed to exist right now. |
| 2782 | */ |
| 2783 | getSegmentMatches(words: string[], limit: number = 6): SegmentMatch[] { |
| 2784 | if (words.length === 0) return []; |
| 2785 | // Variant → original word (plural folding), for coverage accounting. |
| 2786 | const variantToWord = new Map<string, string>(); |
| 2787 | for (const word of words) { |
| 2788 | for (const variant of segmentLookupVariants(word)) { |
| 2789 | if (!variantToWord.has(variant)) variantToWord.set(variant, word); |
| 2790 | } |
| 2791 | } |
| 2792 | const variants = [...variantToWord.keys()]; |
| 2793 | |
| 2794 | // Tier A: co-occurrence. The SQL folds variants back to their original |
| 2795 | // word (#1146), so minWords=2 means two distinct PROMPT WORDS — a name |
| 2796 | // matching both `service` and `services` can't tie with (or crowd past |
| 2797 | // the LIMIT) a genuine two-word match. The JS re-check below recomputes |
| 2798 | // the fold from live segments as the honesty layer. |
| 2799 | const variantPairs = [...variantToWord.entries()].map(([segment, word]) => ({ segment, word })); |
| 2800 | const candidates: Array<{ name: string; matchedWords: Set<string> }> = []; |
| 2801 | for (const hit of this.queries.getSegmentCoOccurrence(variantPairs, 2, 24)) { |
| 2802 | const matched = this.wordsMatchingName(hit.name, variantToWord); |
| 2803 | if (matched.size >= 2) candidates.push({ name: hit.name, matchedWords: matched }); |
| 2804 | } |
| 2805 | |
| 2806 | // Tier B: single rare word. Only when co-occurrence found nothing — a |
| 2807 | // co-occurring name is categorically stronger evidence — and under |
| 2808 | // stricter rules, because one word is thin: the word must be ≥5 chars |
| 2809 | // (measured FPs: "this", "typo"); the segment must appear in AT LEAST TWO |
| 2810 | // names (a concept the codebase is about clusters across names — |
| 2811 | // CheckoutService/CheckoutController — while a prose coincidence is a |
| 2812 | // singleton: measured FP "deploy to PRODUCTION" → the one name |
| 2813 | // matchesNonProductionDir); and the candidate name must have ≥2 segments |
| 2814 | // (a bare common verb matching a bare function name — "write" → `write` — |
| 2815 | // is prose coincidence, not the user naming a symbol). |
| 2816 | if (candidates.length === 0) { |
| 2817 | const singleWordVariants = variants.filter((v) => variantToWord.get(v)!.length >= 5); |
| 2818 | const counts = this.queries.getSegmentNameCounts(singleWordVariants); |
| 2819 | const rare = [...counts.entries()] |
| 2820 | .filter(([, n]) => n >= 2 && n <= CodeGraph.SEGMENT_RARITY_CEILING) |
| 2821 | .sort((a, b) => a[1] - b[1]) |
| 2822 | .slice(0, 2); |
| 2823 | for (const [variant] of rare) { |
| 2824 | const word = variantToWord.get(variant)!; |
| 2825 | for (const name of this.queries.getNamesForSegment(variant, 12)) { |
| 2826 | if (splitIdentifierSegments(name).length < 2) continue; |
| 2827 | candidates.push({ name, matchedWords: new Set([word]) }); |
| 2828 | } |
| 2829 | } |
| 2830 | } |
| 2831 | |
| 2832 | // Verify against nodes (the honesty gate) and pick a representative |
| 2833 | // definition per name. A name whose only nodes are file/import kind has |
| 2834 | // no real definition to point at — surfacing the import statement instead |
| 2835 | // reads as a matched symbol but isn't one (#1144) — so it's skipped, the |
| 2836 | // same way an orphaned vocab row is. (Import names no longer enter the |
| 2837 | // vocab at write time, but rows written before that exclusion persist |
| 2838 | // until the next full index.) |
| 2839 | const out: SegmentMatch[] = []; |
| 2840 | const seen = new Set<string>(); |
no test coverage detected