@jefuriiij/synthra 0.2.0 → 0.2.1

package/CHANGELOG.md

@@ -7,6 +7,22 @@ For older versions, see [GitHub Releases](https://github.com/jefuriiij/synthra/r
 
 ---
 
+## [0.2.1] — 2026-06-06
+
+### Changed
+
+- **Keyword retrieval is now IDF-weighted (BM25's term-rarity component).** A
+  query token that's rare across the repo counts for more than a common one, so
+  on a multi-term query the files matching the *specific* terms rank above those
+  matching generic ones — instead of every keyword match counting the same. The
+  weighting is normalized to the query's mean IDF, so a typical match scores the
+  same as before: overall ranking magnitude — and the confidence / Moat gating
+  that depends on it — is unchanged. Purely an in-repo ranking refinement, no API
+  or data-model change. (TF-saturation / length-norm parts of full BM25 don't
+  apply to the deduped top-N keyword representation.)
+
+---
+
 ## [0.2.0] — 2026-06-06
 
 ### Added

package/dist/cli/index.js

@@ -18,7 +18,7 @@ var init_package = __esm({
   "package.json"() {
     package_default = {
       name: "@jefuriiij/synthra",
-      version: "0.2.0",
+      version: "0.2.1",
       publishConfig: {
         access: "public"
       },
@@ -3880,6 +3880,7 @@ import { appendFile as appendFile3, mkdir as mkdir9 } from "fs/promises";
 import { dirname as dirname10 } from "path";
 
 // src/graph/rank.ts
+var KW_BASE_WEIGHT = 2;
 var USAGE_BOOST_CAP_DEFAULT = 4;
 function usageBoostCap() {
   const env = Number(process.env.SYN_LEARN_BOOST_CAP);
@@ -3969,14 +3970,41 @@ function scoreFiles(inputs) {
   const importsFrom = indexImportEdges(inputs.graph);
   const seeds = new Set(inputs.sessionKnownPaths ?? []);
   for (const p of inputs.recentlyEditedPaths ?? []) seeds.add(p);
+  const corpusSize = inputs.candidates.length;
+  const queryDf = /* @__PURE__ */ new Map();
+  for (const f of inputs.candidates) {
+    for (const kw of f.keywords) {
+      if (qTokens.has(kw)) queryDf.set(kw, (queryDf.get(kw) ?? 0) + 1);
+    }
+  }
+  const idf = (token) => {
+    const n = queryDf.get(token) ?? 0;
+    if (n <= 0) return 0;
+    return Math.log(1 + (corpusSize - n + 0.5) / (n + 0.5));
+  };
+  let idfSum = 0;
+  let idfCount = 0;
+  for (const t of qTokens) {
+    const v = idf(t);
+    if (v > 0) {
+      idfSum += v;
+      idfCount += 1;
+    }
+  }
+  const refIdf = idfCount > 0 ? idfSum / idfCount : 1;
   const scored = [];
   for (const file of inputs.candidates) {
     const reasons = [];
     let score2 = 0;
     let kwHits = 0;
-    for (const kw of file.keywords) if (qTokens.has(kw)) kwHits += 1;
+    let kwScore = 0;
+    for (const kw of file.keywords) {
+      if (!qTokens.has(kw)) continue;
+      kwHits += 1;
+      kwScore += KW_BASE_WEIGHT * (idf(kw) / refIdf);
+    }
     if (kwHits) {
-      score2 += kwHits * 2;
+      score2 += kwScore;
       reasons.push(`kw=${kwHits}`);
     }
     const symbols = symbolsByFile.get(file.path) ?? [];