code-graph-context 2.10.4 → 2.11.0

package/README.md

@@ -189,8 +189,9 @@ If you prefer to edit the config files directly:
 | `NEO4J_URI` | No | `bolt://localhost:7687` | Neo4j connection URI |
 | `NEO4J_USER` | No | `neo4j` | Neo4j username |
 | `NEO4J_PASSWORD` | No | `PASSWORD` | Neo4j password |
-| `EMBEDDING_MODEL` | No | `Qodo/Qodo-Embed-1-1.5B` | Local embedding model (see [Embedding Configuration](#embedding-configuration)) |
+| `EMBEDDING_MODEL` | No | `Qwen/Qwen3-Embedding-0.6B` | Local embedding model (see [Embedding Configuration](#embedding-configuration)) |
 | `EMBEDDING_SIDECAR_PORT` | No | `8787` | Port for local embedding server |
+| `EMBEDDING_HALF_PRECISION` | No | `false` | Set `true` for float16 (uses ~0.5x memory) |
 | `OPENAI_ENABLED` | No | `false` | Set `true` to use OpenAI instead of local |
 | `OPENAI_API_KEY` | No* | - | Required when `OPENAI_ENABLED=true` |
 
@@ -535,20 +536,25 @@ This enables queries like "find all hooks that use context" while maintaining AS
 
 Local embeddings are the default — **no API key needed**. The Python sidecar starts automatically on first use and runs a local model for high-quality code embeddings.
 
+The sidecar uses **float16 (half precision)** by default, which halves memory usage with no meaningful quality loss. It also auto-shuts down after 3 minutes of inactivity to free memory, and restarts lazily when needed (~15-20s).
+
+> **Half precision mode:** To reduce memory usage at the cost of some accuracy, set `EMBEDDING_HALF_PRECISION=true`.
+
 ### Available Models
 
 Set via the `EMBEDDING_MODEL` environment variable:
 
-| Model | Dimensions | RAM | Quality | Best For |
+| Model | Dimensions | RAM (fp16) | Quality | Best For |
 |-------|-----------|-----|---------|----------|
-| `Qodo/Qodo-Embed-1-1.5B` (default) | 1536 | ~9 GB | Best | Machines with 32+ GB RAM |
-| `BAAI/bge-base-en-v1.5` | 768 | ~500 MB | Good | General purpose, low RAM |
-| `sentence-transformers/all-MiniLM-L6-v2` | 384 | ~200 MB | OK | Minimal RAM, fast |
-| `nomic-ai/nomic-embed-text-v1.5` | 768 | ~600 MB | Good | Code + prose mixed |
-| `sentence-transformers/all-mpnet-base-v2` | 768 | ~500 MB | Good | Balanced quality/speed |
-| `BAAI/bge-small-en-v1.5` | 384 | ~130 MB | OK | Smallest footprint |
-
-**Example:** Use a lightweight model on a 16GB machine:
+| `Qwen/Qwen3-Embedding-0.6B` (default) | 1024 | ~1.2 GB | Best | Default, code-aware, MTEB-Code #1 |
+| `Qodo/Qodo-Embed-1-1.5B` | 1536 | ~4.5 GB | Great | Machines with 32+ GB RAM |
+| `BAAI/bge-base-en-v1.5` | 768 | ~250 MB | Good | General purpose, low RAM |
+| `sentence-transformers/all-MiniLM-L6-v2` | 384 | ~100 MB | OK | Minimal RAM, fast |
+| `nomic-ai/nomic-embed-text-v1.5` | 768 | ~300 MB | Good | Code + prose mixed |
+| `sentence-transformers/all-mpnet-base-v2` | 768 | ~250 MB | Good | Balanced quality/speed |
+| `BAAI/bge-small-en-v1.5` | 384 | ~65 MB | OK | Smallest footprint |
+
+**Example:** Use a lightweight model on a low-memory machine:
 ```bash
 claude mcp add --scope user code-graph-context \
   -e EMBEDDING_MODEL=BAAI/bge-base-en-v1.5 \

package/dist/cli/cli.js

@@ -251,7 +251,7 @@ const setupSidecar = async () => {
         return;
     }
     // Pre-download the embedding model so first real use is fast
-    const modelName = process.env.EMBEDDING_MODEL ?? 'Qodo/Qodo-Embed-1-1.5B';
+    const modelName = process.env.EMBEDDING_MODEL ?? 'Qwen/Qwen3-Embedding-0.6B';
     await preDownloadModel(sidecarDir, python, modelName);
 };
 /**

package/dist/core/embeddings/embedding-sidecar.js

@@ -12,9 +12,10 @@ const __dirname = dirname(__filename);
 const DEFAULT_CONFIG = {
     port: parseInt(process.env.EMBEDDING_SIDECAR_PORT ?? '', 10) || 8787,
     host: '127.0.0.1',
-    model: process.env.EMBEDDING_MODEL ?? 'Qodo/Qodo-Embed-1-1.5B',
+    model: process.env.EMBEDDING_MODEL ?? 'Qwen/Qwen3-Embedding-0.6B',
     startupTimeoutMs: 120_000, // 2 min — first run downloads the model
     requestTimeoutMs: 60_000,
+    idleTimeoutMs: 180_000, // 3 min — auto-shutdown after no requests
 };
 export class EmbeddingSidecar {
     process = null;
@@ -23,6 +24,7 @@ export class EmbeddingSidecar {
     _dimensions = null;
     stopping = false;
     _exitHandler = null;
+    _idleTimer = null;
     constructor(config = {}) {
         this.config = { ...DEFAULT_CONFIG, ...config };
     }
@@ -205,6 +207,7 @@ export class EmbeddingSidecar {
             const data = (await res.json());
             if (data.dimensions)
                 this._dimensions = data.dimensions;
+            this.resetIdleTimer();
             return data.embeddings;
         }
         catch (err) {
@@ -251,7 +254,21 @@ export class EmbeddingSidecar {
         }
         this.cleanup();
     }
+    resetIdleTimer() {
+        if (this._idleTimer)
+            clearTimeout(this._idleTimer);
+        this._idleTimer = setTimeout(() => {
+            console.error(`[embedding-sidecar] Idle for ${this.config.idleTimeoutMs / 1000}s, shutting down to free memory`);
+            this.stop();
+        }, this.config.idleTimeoutMs);
+        // Don't let the timer prevent Node from exiting
+        this._idleTimer.unref();
+    }
     cleanup() {
+        if (this._idleTimer) {
+            clearTimeout(this._idleTimer);
+            this._idleTimer = null;
+        }
         if (this._exitHandler) {
             process.removeListener('exit', this._exitHandler);
             this._exitHandler = null;

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

@@ -5,7 +5,7 @@
  * and get the right implementation based on OPENAI_ENABLED.
  *
  *   OPENAI_ENABLED=true  → OpenAI text-embedding-3-large (requires OPENAI_API_KEY)
- *   default              → Local Python sidecar with Qodo-Embed-1-1.5B
+ *   default              → Local Python sidecar with Qwen3-Embedding-0.6B
  */
 import { LocalEmbeddingsService } from './local-embeddings.service.js';
 import { OpenAIEmbeddingsService } from './openai-embeddings.service.js';
@@ -24,6 +24,7 @@ export const EMBEDDING_DIMENSIONS = {
     'text-embedding-3-large': 3072,
     'text-embedding-3-small': 1536,
     // Local models (via sidecar)
+    'Qwen/Qwen3-Embedding-0.6B': 1024,
     'Qodo/Qodo-Embed-1-1.5B': 1536,
     'sentence-transformers/all-MiniLM-L6-v2': 384,
     'sentence-transformers/all-mpnet-base-v2': 768,
@@ -45,7 +46,7 @@ export const getEmbeddingDimensions = () => {
         const model = process.env.OPENAI_EMBEDDING_MODEL ?? 'text-embedding-3-large';
         return EMBEDDING_DIMENSIONS[model] ?? 3072;
     }
-    const model = process.env.EMBEDDING_MODEL ?? 'Qodo/Qodo-Embed-1-1.5B';
+    const model = process.env.EMBEDDING_MODEL ?? 'Qwen/Qwen3-Embedding-0.6B';
     return EMBEDDING_DIMENSIONS[model] ?? 1536;
 };
 /**

package/dist/mcp/tools/parse-typescript-project.tool.js

@@ -10,6 +10,7 @@ import { fileURLToPath } from 'url';
 import { Worker } from 'worker_threads';
 import { z } from 'zod';
 import { CORE_TYPESCRIPT_SCHEMA } from '../../core/config/schema.js';
+import { stopEmbeddingSidecar } from '../../core/embeddings/embedding-sidecar.js';
 import { EmbeddingsService } from '../../core/embeddings/embeddings.service.js';
 import { ParserFactory } from '../../core/parsers/parser-factory.js';
 import { detectChangedFiles } from '../../core/utils/file-change-detection.js';
@@ -165,6 +166,7 @@ export const createParseTypescriptProjectTool = (server) => {
                     const job = jobManager.getJob(jobId);
                     if (job && job.status === 'running') {
                         jobManager.failJob(jobId, `Worker timed out after ${PARSING.workerTimeoutMs / 60000} minutes`);
+                        await stopEmbeddingSidecar();
                         await terminateWorker('timeout');
                     }
                 }, PARSING.workerTimeoutMs);
@@ -183,6 +185,7 @@ export const createParseTypescriptProjectTool = (server) => {
                         clearTimeout(timeoutId);
                         jobManager.failJob(jobId, msg.error);
                         debugLog('Async parsing failed', { jobId, error: msg.error });
+                        stopEmbeddingSidecar();
                         terminateWorker('error');
                     }
                 });
@@ -191,6 +194,7 @@ export const createParseTypescriptProjectTool = (server) => {
                     clearTimeout(timeoutId);
                     jobManager.failJob(jobId, err.message ?? String(err));
                     console.error('Worker thread error:', err);
+                    stopEmbeddingSidecar();
                     terminateWorker('worker-error');
                 });
                 // Handle worker exit
@@ -371,12 +375,16 @@ export const createParseTypescriptProjectTool = (server) => {
                     edgeCount: 0,
                 });
                 await debugLog('Project status updated to failed', { projectId: finalProjectId });
+                // Stop sidecar to free memory (restarts lazily on next embed request)
+                await stopEmbeddingSidecar();
                 return createSuccessResponse(formatParsePartialSuccess(nodes.length, edges.length, outputPath, neo4jError.message));
             }
         }
         catch (error) {
             console.error('Parse tool error:', error);
             await debugLog('Parse tool error', { projectPath, tsconfigPath, error });
+            // Stop sidecar to free memory (restarts lazily on next embed request)
+            await stopEmbeddingSidecar();
             return createErrorResponse(error);
         }
     });

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "code-graph-context",
-  "version": "2.10.4",
+  "version": "2.11.0",
   "description": "MCP server that builds code graphs to provide rich context to LLMs",
   "type": "module",
   "homepage": "https://github.com/drewdrewH/code-graph-context#readme",

package/sidecar/embedding_server.py

@@ -1,6 +1,6 @@
 """
 Local embedding server for code-graph-context.
-Uses Qodo-Embed-1-1.5B for high-quality code embeddings without OpenAI dependency.
+Uses Qwen3-Embedding-0.6B for high-quality code embeddings without OpenAI dependency.
 Runs as a sidecar process managed by the Node.js MCP server.
 """
 
@@ -27,7 +27,7 @@ logger.info(f"Sidecar process starting (pid={os.getpid()})")
 app = FastAPI(title="code-graph-context embedding sidecar")
 
 model = None
-model_name = os.environ.get("EMBEDDING_MODEL", "Qodo/Qodo-Embed-1-1.5B")
+model_name = os.environ.get("EMBEDDING_MODEL", "Qwen/Qwen3-Embedding-0.6B")
 
 
 class EmbedRequest(BaseModel):
@@ -52,8 +52,14 @@ def load_model():
         logger.info(f"Loading {model_name} on {device}...")
         logger.info(f"PyTorch version: {torch.__version__}, MPS available: {torch.backends.mps.is_available()}")
 
+        use_half = os.environ.get("EMBEDDING_HALF_PRECISION", "").lower() == "true"
         model = SentenceTransformer(model_name, device=device)
-        logger.info(f"Model loaded into memory, running warmup...")
+        if use_half:
+            model.half()
+            logger.info(f"Model loaded in float16 (half precision)")
+        else:
+            logger.info(f"Model loaded in float32 (full precision)")
+        logger.info(f"Running warmup...")
 
         # Warm up with a test embedding
         with torch.no_grad():

package/sidecar/requirements.txt

@@ -3,3 +3,4 @@ uvicorn>=0.24.0
 sentence-transformers>=3.0.0
 torch>=2.0.0
 pydantic>=2.0.0
+transformers>=4.51.0