embarsy-qdrant-mcp 0.1.2 → 0.1.3

package/dist/chunker.js

@@ -39,13 +39,26 @@ export function chunkFile(content, cfg) {
             if (nl > pos + Math.floor(maxChars / 2))
                 end = nl;
         }
+        end = snapToCodePoint(content, end);
         const text = content.slice(pos, end);
         if (text.trim().length > 0) {
             chunks.push({ startLine: lineAt(pos), endLine: lineAt(Math.max(pos, end - 1)), text });
         }
         if (end >= content.length)
             break;
-        pos = Math.max(pos + 1, end - overlap);
+        pos = snapToCodePoint(content, Math.max(pos + 1, end - overlap));
     }
     return chunks;
 }
+/** A boundary that lands between the two halves of a surrogate pair would cut an emoji
+ *  in half; the resulting lone surrogate is invalid Unicode that embedding backends
+ *  reject (Python's UTF-8 encoder raises on it, turning the request into a 500).
+ *  If the index points at a low surrogate, step back one unit to keep the pair whole. */
+function snapToCodePoint(s, index) {
+    if (index > 0 && index < s.length) {
+        const code = s.charCodeAt(index);
+        if (code >= 0xdc00 && code <= 0xdfff)
+            return index - 1;
+    }
+    return index;
+}

package/dist/embeddings.js

@@ -7,10 +7,17 @@ import { requestJSON } from "./http.js";
 /** Absolute per-input cap (defense in depth on top of chunking) so no single text can overflow
  *  the model's context and crash the embedding backend, whatever produced it. */
 const MAX_EMBED_CHARS = 8000;
+/** Lone surrogates (an emoji cut in half by any UTF-16 slice) are invalid Unicode: the
+ *  Embarsy API's UTF-8 re-encode for Ollama rejects them, failing the whole batch with
+ *  a 500. Replace them with U+FFFD so one broken character can never sink an index run. */
+const LONE_SURROGATE = /[\uD800-\uDBFF](?![\uDC00-\uDFFF])|(?<![\uD800-\uDBFF])[\uDC00-\uDFFF]/g;
+function wellFormed(text) {
+    return text.replace(LONE_SURROGATE, "�");
+}
 export async function embedBatch(texts, cfg) {
     if (texts.length === 0)
         return [];
-    const input = texts.map((t) => (t.length > MAX_EMBED_CHARS ? t.slice(0, MAX_EMBED_CHARS) : t));
+    const input = texts.map((t) => wellFormed(t.length > MAX_EMBED_CHARS ? t.slice(0, MAX_EMBED_CHARS) : t));
     const json = await requestJSON(`${cfg.openaiBaseUrl}/embeddings`, {
         method: "POST",
         headers: {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "embarsy-qdrant-mcp",
-  "version": "0.1.2",
+  "version": "0.1.3",
   "description": "Codebase-indexing MCP server for Embarsy — semantic code search over a local Qdrant + OpenAI-compatible embeddings stack. Routes embeddings and Qdrant through Embarsy's proxy so auth and Monitoring counters just work.",
   "type": "module",
   "bin": {