code-graph-context 2.10.5 → 2.11.1

package/README.md

@@ -189,9 +189,10 @@ 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_FULL_PRECISION` | No | `false` | Set `true` for float32 (uses ~2x memory) |
+| `EMBEDDING_DEVICE` | No | `cpu` | Device for embeddings (`cpu` or `mps`). CPU is default to avoid MPS memory bloat |
+| `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` |
 
@@ -536,9 +537,11 @@ 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).
+The sidecar runs on **CPU by default** to avoid MPS memory pool bloat on Apple Silicon (MPS can pre-allocate 10+ GB even for small models). CPU is fast enough for models under 1B params. It also auto-shuts down after 3 minutes of inactivity to free memory, and restarts lazily when needed (~15-20s).
 
-> **Full precision mode:** If you have 32+ GB RAM and want float32, set `EMBEDDING_FULL_PRECISION=true`.
+> **GPU acceleration:** Set `EMBEDDING_DEVICE=mps` to use Apple Silicon GPU for larger models (1B+ params). Only recommended on machines with 32+ GB RAM.
+>
+> **Half precision:** Set `EMBEDDING_HALF_PRECISION=true` to load the model in float16, roughly halving memory usage.
 
 ### Available Models
 
@@ -546,7 +549,8 @@ Set via the `EMBEDDING_MODEL` environment variable:
 
 | Model | Dimensions | RAM (fp16) | Quality | Best For |
 |-------|-----------|-----|---------|----------|
-| `Qodo/Qodo-Embed-1-1.5B` (default) | 1536 | ~4.5 GB | Best | Default, works on 16GB machines |
+| `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 |

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

@@ -12,7 +12,7 @@ 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

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/package.json

@@ -1,6 +1,6 @@
 {
   "name": "code-graph-context",
-  "version": "2.10.5",
+  "version": "2.11.1",
   "description": "MCP server that builds code graphs to provide rich context to LLMs",
   "type": "module",
   "homepage": "https://github.com/drewdrewH/code-graph-context#readme",

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):
@@ -48,16 +48,23 @@ def load_model():
         import torch
         from sentence_transformers import SentenceTransformer
 
-        device = "mps" if torch.backends.mps.is_available() else "cpu"
+        # Use CPU by default — MPS pre-allocates a massive memory pool (10+ GB)
+        # that bloats small models. CPU on Apple Silicon is fast enough for <1B models.
+        # Set EMBEDDING_DEVICE=mps to force GPU if needed for large models.
+        device_override = os.environ.get("EMBEDDING_DEVICE", "").lower()
+        if device_override:
+            device = device_override
+        else:
+            device = "cpu"
         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_FULL_PRECISION", "").lower() != "true"
-        model = SentenceTransformer(model_name, device=device)
+        use_half = os.environ.get("EMBEDDING_HALF_PRECISION", "").lower() == "true"
         if use_half:
-            model.half()
+            model = SentenceTransformer(model_name, device=device, model_kwargs={"torch_dtype": "float16"})
             logger.info(f"Model loaded in float16 (half precision)")
         else:
+            model = SentenceTransformer(model_name, device=device)
             logger.info(f"Model loaded in float32 (full precision)")
         logger.info(f"Running warmup...")
 
@@ -114,7 +121,6 @@ async def embed(req: EmbedRequest):
 def _encode_with_oom_fallback(texts: list[str], batch_size: int) -> list[list[float]]:
     """
     Encode texts, falling back to CPU if MPS runs out of memory.
-    Also retries with smaller batch sizes before giving up.
     """
     import torch
 
@@ -126,28 +132,18 @@ def _encode_with_oom_fallback(texts: list[str], batch_size: int) -> list[list[fl
                 show_progress_bar=False,
                 normalize_embeddings=True,
             )
-        # Free intermediate tensors after each request
-        if hasattr(torch.mps, "empty_cache"):
-            torch.mps.empty_cache()
         return result.tolist()
-    except (torch.mps.OutOfMemoryError, RuntimeError) as e:
+    except (RuntimeError,) as e:
         if "out of memory" not in str(e).lower():
             raise
 
-        logger.warning(f"MPS OOM with batch_size={batch_size}, len={len(texts)}. Falling back to CPU.")
-
-        # Free MPS memory
-        if hasattr(torch.mps, "empty_cache"):
-            torch.mps.empty_cache()
+        logger.warning(f"OOM with batch_size={batch_size}, len={len(texts)}. Falling back to CPU.")
         gc.collect()
 
-        # Fall back to CPU for this request
         original_device = model.device
         model.to("cpu")
-        logger.info("Model moved to CPU for fallback encoding")
 
         try:
-            # Use smaller batches on CPU
             cpu_batch = min(batch_size, 4)
             with torch.no_grad():
                 result = model.encode(
@@ -159,12 +155,10 @@ def _encode_with_oom_fallback(texts: list[str], batch_size: int) -> list[list[fl
             logger.info(f"CPU fallback encoding complete ({len(texts)} texts)")
             return result.tolist()
         finally:
-            # Move back to MPS for future requests
             try:
                 model.to(original_device)
-                logger.info(f"Model moved back to {original_device}")
             except Exception:
-                logger.warning("Could not move model back to MPS, staying on CPU")
+                logger.warning("Could not move model back, staying on CPU")
 
 
 def handle_signal(sig, _frame):

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