@oomkapwn/enquire-mcp 2.7.0 → 2.9.0

package/dist/tools.js

@@ -2366,7 +2366,8 @@ export async function embeddingsSearch(vault, args, embedFile) {
             snippet: h.text_preview.slice(0, 240),
             chunk_index: h.chunk_index,
             line_start: h.line_start,
-            line_end: h.line_end
+            line_end: h.line_end,
+            kind: h.kind
         }));
         return { query: args.query, method: "embeddings-cosine", model: model.alias, total_chunks: total, matches };
     }
@@ -2411,7 +2412,8 @@ export async function searchHybrid(vault, args, ctx) {
                     snippet: h.snippet,
                     chunk_index: h.chunk_index,
                     line_start: h.line_start,
-                    line_end: h.line_end
+                    line_end: h.line_end,
+                    kind: h.kind
                 }));
             }
             else {
@@ -2425,7 +2427,8 @@ export async function searchHybrid(vault, args, ctx) {
                             snippet: h.snippet,
                             chunk_index: h.chunk_index,
                             line_start: h.line_start,
-                            line_end: h.line_end
+                            line_end: h.line_end,
+                            kind: h.kind
                         });
                     }
                 });
@@ -2436,7 +2439,8 @@ export async function searchHybrid(vault, args, ctx) {
                     snippet: b.snippet,
                     chunk_index: b.chunk_index,
                     line_start: b.line_start,
-                    line_end: b.line_end
+                    line_end: b.line_end,
+                    kind: b.kind
                 }));
                 // Re-sort to ensure 1-based ranks are consecutive after dedup.
                 bm25Ranked.sort((a, b) => a.rank - b.rank);
@@ -2499,7 +2503,8 @@ export async function searchHybrid(vault, args, ctx) {
                     snippet: m.snippet,
                     chunk_index: m.chunk_index,
                     line_start: m.line_start,
-                    line_end: m.line_end
+                    line_end: m.line_end,
+                    kind: m.kind
                 }));
             }
             else {
@@ -2513,7 +2518,8 @@ export async function searchHybrid(vault, args, ctx) {
                             snippet: m.snippet,
                             chunk_index: m.chunk_index,
                             line_start: m.line_start,
-                            line_end: m.line_end
+                            line_end: m.line_end,
+                            kind: m.kind
                         });
                     }
                 });
@@ -2524,7 +2530,8 @@ export async function searchHybrid(vault, args, ctx) {
                     snippet: b.snippet,
                     chunk_index: b.chunk_index,
                     line_start: b.line_start,
-                    line_end: b.line_end
+                    line_end: b.line_end,
+                    kind: b.kind
                 }));
                 embedRanked.sort((a, b) => a.rank - b.rank);
                 for (let i = 0; i < embedRanked.length; i++) {
@@ -2605,6 +2612,71 @@ export async function searchHybrid(vault, args, ctx) {
     const bm25Map = new Map(bm25Ranked.map((h) => [h.id, h]));
     const tfidfMap = new Map(tfidfRanked.map((h) => [h.id, h]));
     const embedMap = new Map(embedRanked.map((h) => [h.id, h]));
+    // ─── v2.9.0: Cross-encoder reranking (post-RRF, post-graph-boost) ────────
+    // Take the top-N fused candidates, score each (query, snippet) pair with a
+    // BGE-style cross-encoder, and re-sort. Cross-encoder is far more accurate
+    // than bi-encoder cosine for relevance ranking — it sees query+document
+    // interaction directly. ~30-50ms per query overhead on M1 CPU at N=50.
+    //
+    // Failures are caught and surfaced as `signal_errors.reranker` so a model
+    // load problem doesn't poison the whole search response. The fused order
+    // (RRF + graph-boost) is preserved if reranking fails.
+    let rerankerScores = null;
+    if ((ctx.reranker || ctx.rerankerOverride) && fused.length > 0) {
+        const topN = ctx.reranker?.topN ?? 50;
+        const rerankBatch = fused.slice(0, topN);
+        try {
+            // Prefer the test-injected reranker when present; otherwise lazy-load.
+            let reranker;
+            if (ctx.rerankerOverride) {
+                reranker = ctx.rerankerOverride;
+            }
+            else {
+                const { loadReranker } = await import("./embeddings.js");
+                reranker = await loadReranker(ctx.reranker?.alias);
+            }
+            // For each candidate, find the best snippet (BM25 > embeddings > TF-IDF)
+            // and pair it with the query. Empty-snippet candidates go to the bottom
+            // by getting a -Infinity score (sort below scored candidates).
+            const passages = rerankBatch.map((f) => {
+                const bm = bm25Map.get(f.id);
+                const emb = embedMap.get(f.id);
+                const tf = tfidfMap.get(f.id);
+                const snippet = bm?.snippet ?? emb?.snippet ?? tf?.snippet ?? "";
+                // Strip FTS5 «…» highlight markers — they're cosmetic and the
+                // reranker should see clean prose. Limit to ~600 chars to stay
+                // safely under the model's 512-token budget (rough char/token ratio
+                // varies by language; 600 chars ≈ 200 tokens for English / Cyrillic
+                // per the multilingual model's tokenizer, well under 512).
+                return snippet.replace(/[«»]/g, "").slice(0, 600);
+            });
+            const scores = await reranker.score(args.query, passages);
+            rerankerScores = new Map();
+            for (let i = 0; i < rerankBatch.length; i++) {
+                const f = rerankBatch[i];
+                const s = scores[i];
+                if (f && typeof s === "number")
+                    rerankerScores.set(f.id, s);
+            }
+            // Sort the top-N by reranker score; everything below top-N keeps RRF
+            // order. We do this by re-ordering fused[0..topN] in place.
+            const reordered = [...rerankBatch].sort((a, b) => {
+                const sa = rerankerScores?.get(a.id) ?? -Infinity;
+                const sb = rerankerScores?.get(b.id) ?? -Infinity;
+                return sb - sa;
+            });
+            for (let i = 0; i < reordered.length; i++) {
+                fused[i] = reordered[i];
+            }
+        }
+        catch (err) {
+            const msg = err instanceof Error ? err.message : String(err);
+            // Add to signalErrors so it surfaces in the response. Reranker is not
+            // a "signal" per se but the existing dict is the right home.
+            signalErrors.reranker = msg;
+            process.stderr.write(`obsidian_search: reranker failed — ${msg}\n`);
+        }
+    }
     const matches = [];
     for (const f of fused) {
         const numSignals = Object.keys(f.per_signal).length;
@@ -2642,15 +2714,29 @@ export async function searchHybrid(vault, args, ctx) {
                     chunkFromId = parsed;
             }
         }
+        // v2.8.0: derive content-source kind. BM25 / embeddings hits carry it
+        // explicitly; TF-IDF doesn't (it only runs over markdown). Either
+        // ranker reporting "pdf" wins; otherwise fall back to "md".
+        const kind = bm?.kind === "pdf" || emb?.kind === "pdf" ? "pdf" : "md";
+        // For PDFs, the title is best derived from the filename without
+        // `.md`-stripping (PDFs don't have that extension); use the .pdf-stripped
+        // form so titles read naturally in agent output.
+        const baseName = path.basename(pathPart);
+        const title = kind === "pdf" ? baseName.replace(/\.pdf$/i, "") : stripMd(baseName);
+        const rerankerScore = rerankerScores?.get(f.id);
         matches.push({
             path: pathPart,
-            title: stripMd(path.basename(pathPart)),
+            title,
             score: Math.round(f.score * 100000) / 100000,
             snippet: bestEvidence?.snippet ?? "",
             chunk_index: chunkFromId ?? bm?.chunk_index ?? emb?.chunk_index,
             line_start: bm?.line_start ?? emb?.line_start,
             line_end: bm?.line_end ?? emb?.line_end,
-            per_signal: perSignal
+            kind,
+            per_signal: perSignal,
+            ...(typeof rerankerScore === "number" && Number.isFinite(rerankerScore)
+                ? { reranker_score: Math.round(rerankerScore * 100000) / 100000 }
+                : {})
         });
         if (matches.length >= limit)
             break;