From e34b78b4f1bc7aac17dc277203cdcadaa16e8a3c Mon Sep 17 00:00:00 2001
From: GautamSharma99 <gautamsharma99067@gmail.com>
Date: Sat, 11 Jul 2026 21:57:22 +0530
Subject: [PATCH] fix: preserve strong BM25-only matches

---
 .../graph-context/graph-rank.ts               |  1 +
 scripts/check-bm25-tokenizer.js               | 37 +++++++++++++++++++
 2 files changed, 38 insertions(+)

diff --git a/libs/knowledge-graph/graph-context/graph-rank.ts b/libs/knowledge-graph/graph-context/graph-rank.ts
index 15318db..8862e49 100644
--- a/libs/knowledge-graph/graph-context/graph-rank.ts
+++ b/libs/knowledge-graph/graph-context/graph-rank.ts
@@ -45,6 +45,7 @@ function applyPostGraphThreshold(
 ): RankedContextDocument[] {
   return ranked.filter((document) =>
     document.signals.semantic_score >= config.ranking.semanticThreshold ||
+    document.signals.bm25_score >= config.ranking.semanticThreshold ||
     document.signals.graph_score >= config.ranking.semanticThreshold ||
     document.signals.coherence_score >= config.ranking.semanticThreshold,
   );
diff --git a/scripts/check-bm25-tokenizer.js b/scripts/check-bm25-tokenizer.js
index f880b45..645f763 100644
--- a/scripts/check-bm25-tokenizer.js
+++ b/scripts/check-bm25-tokenizer.js
@@ -3,6 +3,8 @@ import assert from "node:assert/strict";
 const root = new URL("..", import.meta.url);
 const { tokenize, scoreBm25 } = await import(new URL("dist/libs/knowledge-graph/graph-context/bm25.js", root));
 const { graphContextConfig } = await import(new URL("dist/libs/knowledge-graph/graph-context/config.js", root));
+const { applyGraphRanking } = await import(new URL("dist/libs/knowledge-graph/graph-context/graph-rank.js", root));
+const { rankContextDocuments } = await import(new URL("dist/libs/knowledge-graph/graph-context/rank.js", root));
 
 assert.ok(tokenize("handleUserAuth").includes("user"), "camelCase should emit sub-token user");
 assert.ok(tokenize("handleUserAuth").includes("auth"), "camelCase should emit sub-token auth");
@@ -25,4 +27,39 @@ const documents = [{ key: "doc:auth", text: "The handleUserAuth function validat
 const ranked = scoreBm25("user auth validation", documents, graphContextConfig);
 assert.equal(ranked[0]?.id, "doc:auth", "BM25 should match camelCase doc tokens to spaced query terms");
 
+const lexicalClaim = {
+  id: "claim.lexical_match",
+  kind: "fact",
+  text: "The processRareIdentifier function handles this workflow.",
+  truth: "unknown",
+  intent: "unknown",
+};
+const lexicalDocument = {
+  key: "claim:claim.lexical_match",
+  type: "claim",
+  id: lexicalClaim.id,
+  text: lexicalClaim.text,
+  object: lexicalClaim,
+  about: [],
+};
+const lexicalScores = scoreBm25("process rare identifier", [lexicalDocument], graphContextConfig);
+const baseRanked = rankContextDocuments(
+  [lexicalDocument],
+  [{ id: lexicalDocument.key, score: 0, raw_score: 0, rank: 1 }],
+  lexicalScores,
+  graphContextConfig,
+);
+const graphRanked = applyGraphRanking(
+  { claims: baseRanked, components: [], flows: [] },
+  { claims: [lexicalClaim], components: [], flows: [], sources: [], edges: [] },
+  graphContextConfig,
+);
+assert.equal(baseRanked[0]?.signals.bm25_score, 1, "lexical match should have the strongest normalized BM25 score");
+assert.equal(baseRanked[0]?.signals.semantic_score, 0, "regression fixture must not rely on semantic similarity");
+assert.equal(
+  graphRanked.claims[0]?.document.id,
+  lexicalClaim.id,
+  "post-graph filtering should preserve a strong BM25-only match",
+);
+
 console.log("BM25 tokenizer checks passed.");