@vespermcp/mcp-server 1.2.24 → 1.2.26

package/build/metadata/github-source.js

@@ -101,6 +101,18 @@ export class GithubSource {
             qualityWarnings.push("Low star count; may be low-signal");
         if (description.length < 80)
             qualityWarnings.push("Short description; relevance may be weaker");
+        const lowSignalPatterns = [
+            /\bawesome\b/i,
+            /\bresources?\b/i,
+            /\bcurated\b/i,
+            /\blist\b/i,
+            /\bcollection\b/i,
+        ];
+        const lowSignalText = `${fullName} ${description}`;
+        const looksResourceList = lowSignalPatterns.some((rx) => rx.test(lowSignalText));
+        if (looksResourceList) {
+            qualityWarnings.push("Repository appears to be a resource/list collection; relevance may be indirect.");
+        }
         const abstractLength = input.readmeText ? input.readmeText.length : description.length;
         const authorsPresent = !!owner;
         const datePresent = !!updatedAt;
@@ -111,6 +123,19 @@ export class GithubSource {
             datePresent,
             contentDepth,
         });
+        let adjustedQuality01 = quality01;
+        // Calibrate GitHub quality so resource-list repos don't dominate.
+        // Keep penalty moderate; readme-rich/long-form repos still score well.
+        if (looksResourceList) {
+            adjustedQuality01 -= 0.14;
+        }
+        if (!input.readmeText && description.length < 140) {
+            adjustedQuality01 -= 0.08;
+        }
+        if (stars < 50) {
+            adjustedQuality01 -= 0.04;
+        }
+        adjustedQuality01 = Math.max(0.3, Math.min(1.0, adjustedQuality01));
         return {
             id: fullName,
             source: "github",
@@ -133,7 +158,7 @@ export class GithubSource {
                 usage_restrictions: [],
                 warnings: [],
             },
-            quality_score: Math.round(quality01 * 100),
+            quality_score: Math.round(adjustedQuality01 * 100),
             quality_warnings: qualityWarnings,
             download_url: String(repo.html_url || `https://github.com/${fullName}`),
             format: "GIT",

package/build/web/fusion-engine.js

@@ -43,14 +43,70 @@ function tokenize(content) {
         .filter((w) => w.length >= 3);
     return new Set(words);
 }
+function titleTokens(doc) {
+    const mj = doc.metadata_json || {};
+    const raw = typeof mj.title === "string" ? mj.title : "";
+    return tokenize(raw);
+}
+function semanticHintTokens(doc) {
+    const mj = doc.metadata_json || {};
+    const fields = [];
+    if (typeof mj.title === "string")
+        fields.push(mj.title);
+    if (typeof mj.name === "string")
+        fields.push(mj.name);
+    if (typeof mj.description === "string")
+        fields.push(mj.description);
+    if (typeof mj.abstract === "string")
+        fields.push(mj.abstract);
+    if (Array.isArray(mj.tags))
+        fields.push(mj.tags.join(" "));
+    if (Array.isArray(mj.topics))
+        fields.push(mj.topics.join(" "));
+    fields.push(doc.source_url || "");
+    return tokenize(fields.join(" "));
+}
 function isSuspiciousPair(a, b) {
     // semantic fallback should be selective; do cheap prefilter first
+    // Metadata/topic overlap can indicate same object even with very different body lengths.
+    const aHints = semanticHintTokens(a);
+    const bHints = semanticHintTokens(b);
+    if (aHints.size > 0 && bHints.size > 0) {
+        let hInter = 0;
+        for (const t of aHints)
+            if (bHints.has(t))
+                hInter++;
+        const hUnion = aHints.size + bHints.size - hInter;
+        const hJaccard = hUnion > 0 ? hInter / hUnion : 0;
+        if (hJaccard >= 0.2)
+            return true;
+    }
     const aLen = a.content.length;
     const bLen = b.content.length;
     const maxLen = Math.max(aLen, bLen, 1);
     const lenRatio = Math.abs(aLen - bLen) / maxLen;
-    if (lenRatio > 0.45)
+    // Loosened again to allow abstract-vs-summary style comparisons.
+    if (lenRatio > 0.9)
         return false;
+    // Fast path: same normalized title-like prefix often indicates same research object.
+    const aPrefix = a.content.slice(0, 140).toLowerCase().replace(/[^a-z0-9\s]/g, " ").trim();
+    const bPrefix = b.content.slice(0, 140).toLowerCase().replace(/[^a-z0-9\s]/g, " ").trim();
+    if (aPrefix && bPrefix && (aPrefix.includes(bPrefix) || bPrefix.includes(aPrefix))) {
+        return true;
+    }
+    // Cross-source papers often have close titles even if abstracts differ.
+    const aTitle = titleTokens(a);
+    const bTitle = titleTokens(b);
+    if (aTitle.size > 0 && bTitle.size > 0) {
+        let tInter = 0;
+        for (const t of aTitle)
+            if (bTitle.has(t))
+                tInter++;
+        const tUnion = aTitle.size + bTitle.size - tInter;
+        const tJaccard = tUnion > 0 ? tInter / tUnion : 0;
+        if (tJaccard >= 0.25)
+            return true;
+    }
     const aTokens = tokenize(a.content);
     const bTokens = tokenize(b.content);
     if (aTokens.size === 0 || bTokens.size === 0)
@@ -61,7 +117,8 @@ function isSuspiciousPair(a, b) {
             inter++;
     const union = aTokens.size + bTokens.size - inter;
     const jaccard = union > 0 ? inter / union : 0;
-    return jaccard >= 0.18;
+    // Loosened from 0.12 -> 0.08 to let semantic stage inspect more borderline matches.
+    return jaccard >= 0.08;
 }
 function normalizeStars(doc) {
     const mj = doc.metadata_json || {};

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@vespermcp/mcp-server",
-  "version": "1.2.24",
+  "version": "1.2.26",
   "description": "AI-powered dataset discovery, quality analysis, and preparation MCP server with multimodal support (text, image, audio, video)",
   "type": "module",
   "main": "build/index.js",