code-graph-context 2.13.0 → 2.13.1

package/dist/core/embeddings/local-embeddings.service.js

@@ -7,6 +7,10 @@ import { debugLog } from '../../mcp/utils.js';
 import { getEmbeddingSidecar } from './embedding-sidecar.js';
 const BATCH_CONFIG = {
     maxBatchSize: parseInt(process.env.EMBEDDING_BATCH_SIZE ?? '', 10) || 8,
+    // Max texts per HTTP request to the sidecar. Keeps memory bounded when
+    // multiple parallel workers call embedTextsInBatches concurrently.
+    // The sidecar still handles GPU batching internally via batch_size.
+    httpBatchLimit: parseInt(process.env.EMBEDDING_HTTP_BATCH_LIMIT ?? '', 10) || 50,
 };
 export class LocalEmbeddingsService {
     async embedText(text) {
@@ -22,22 +26,30 @@ export class LocalEmbeddingsService {
     async embedTextsInBatches(texts, _batchSize) {
         if (texts.length === 0)
             return [];
-        // GPU batch size controls how many texts the model processes at once (memory-bound).
-        // We send ALL texts in a single HTTP request and let the sidecar handle GPU batching
-        // internally via model.encode(batch_size=N). This eliminates HTTP round-trip overhead.
         const gpuBatchSize = BATCH_CONFIG.maxBatchSize;
-        const gpuBatches = Math.ceil(texts.length / gpuBatchSize);
-        console.error(`[embedding] Sending ${texts.length} texts in 1 request (gpu_batch_size=${gpuBatchSize}, ~${gpuBatches} GPU batches)`);
-        await debugLog('Batch embedding started', { provider: 'local', textCount: texts.length, gpuBatchSize });
+        const httpLimit = BATCH_CONFIG.httpBatchLimit;
+        const httpBatches = Math.ceil(texts.length / httpLimit);
+        const gpuBatchesPerRequest = Math.ceil(httpLimit / gpuBatchSize);
+        console.error(`[embedding] ${texts.length} texts → ${httpBatches} HTTP requests (http_limit=${httpLimit}, gpu_batch_size=${gpuBatchSize}, ~${gpuBatchesPerRequest} GPU batches/req)`);
+        await debugLog('Batch embedding started', { provider: 'local', textCount: texts.length, gpuBatchSize, httpLimit, httpBatches });
         const sidecar = getEmbeddingSidecar();
-        try {
-            const results = await sidecar.embed(texts, gpuBatchSize);
-            return results;
-        }
-        catch (error) {
-            const msg = error instanceof Error ? error.message : String(error);
-            console.error(`[embedding] FAILED (${texts.length} texts, gpuBatchSize=${gpuBatchSize}): ${msg}`);
-            throw error;
+        const allResults = [];
+        for (let i = 0; i < texts.length; i += httpLimit) {
+            const batch = texts.slice(i, i + httpLimit);
+            const batchNum = Math.floor(i / httpLimit) + 1;
+            try {
+                const results = await sidecar.embed(batch, gpuBatchSize);
+                allResults.push(...results);
+                if (httpBatches > 1) {
+                    console.error(`[embedding] HTTP batch ${batchNum}/${httpBatches}: ${batch.length} texts embedded`);
+                }
+            }
+            catch (error) {
+                const msg = error instanceof Error ? error.message : String(error);
+                console.error(`[embedding] FAILED HTTP batch ${batchNum}/${httpBatches} (${batch.length} texts, gpuBatchSize=${gpuBatchSize}): ${msg}`);
+                throw error;
+            }
         }
+        return allResults;
     }
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "code-graph-context",
-  "version": "2.13.0",
+  "version": "2.13.1",
   "description": "MCP server that builds code graphs to provide rich context to LLMs",
   "type": "module",
   "homepage": "https://github.com/drewdrewH/code-graph-context#readme",