@iceinvein/code-intelligence-mcp 1.1.3 → 1.2.0

package/README.md

@@ -17,7 +17,7 @@ Unlike basic text search, this server builds a local knowledge graph to understa
 * **Advanced Hybrid Search**: Combines **Tantivy** (keyword BM25) + **LanceDB** (semantic vector) + **Jina Code embeddings** (768-dim code-specific model) with Reciprocal Rank Fusion (RRF).
 * **Cross-Encoder Reranking**: Always-on ORT-based reranker for precision result ranking.
 * **Smart Context Assembly**: Token-aware budgeting with query-aware truncation that keeps relevant lines within context limits.
-* **On-Device LLM Descriptions**: Automatically generates natural-language descriptions for every symbol using a local **Qwen2.5-Coder-1.5B** model (ONNX Runtime), enriching search with human-readable summaries.
+* **On-Device LLM Descriptions**: Automatically generates natural-language descriptions for every symbol using a local **Qwen2.5-Coder-1.5B** model (llama.cpp with Metal GPU), enriching search with human-readable summaries.
 * **PageRank Scoring**: Graph-based symbol importance scoring that identifies central, heavily-used components.
 * **Learns from Feedback**: Optional learning system that adapts to user selections over time.
 * **Production First**: Multi-layer test detection (file paths, symbol names, and AST-level `#[test]`/`mod tests` analysis) ensures implementation code ranks above test helpers.
@@ -134,18 +134,19 @@ Each client session is bound to its workspace root. The server maintains separat
 
 ### Data Storage
 
-All standalone data lives in `~/.code-intelligence/`:
+Both embedded (stdio) and standalone (HTTP) modes store all data in `~/.code-intelligence/`:
 
 ```text
 ~/.code-intelligence/
-├── server.toml              # Optional config file
-├── models/                  # Shared embedding model (~500MB, loaded once)
-│   └── jina-code-onnx/
+├── server.toml              # Optional config file (standalone only)
+├── models/                  # Shared models (loaded once, shared across repos)
+│   ├── jina-code-onnx/      # Embedding model (~500MB)
+│   └── qwen2.5-coder-1.5b-gguf/  # LLM model (~1.1GB)
 ├── logs/
 │   └── server.log
 └── repos/
     ├── registry.json        # Tracks all known repos
-    ├── a1b2c3d4e5f6a7b8/   # Per-repo data (SHA256 hash of path)
+    ├── a1b2c3d4e5f6a7b8/   # Per-repo data (SHA256 hash of repo path)
     │   ├── code-intelligence.db
     │   ├── tantivy-index/
     │   └── vectors/
@@ -153,6 +154,8 @@ All standalone data lives in `~/.code-intelligence/`:
         └── ...
 ```
 
+The same repo always maps to the same hash regardless of mode, so embedded and standalone can share the same index data.
+
 ### Configuration
 
 Standalone mode is configured via `~/.code-intelligence/server.toml` (created on first run with defaults). Environment variables and CLI flags override TOML settings.
@@ -473,7 +476,7 @@ src/
 │   └── standalone.rs  # Standalone HTTP handler with session routing
 ├── tools/             # Tool definitions (23 MCP tools)
 ├── embeddings/        # Jina Code embedding model wrapper
-├── llm/               # On-device LLM (Qwen2.5-Coder-1.5B via ONNX Runtime)
+├── llm/               # On-device LLM (Qwen2.5-Coder-1.5B via llama.cpp)
 ├── reranker/          # Cross-encoder ORT implementation
 ├── path/              # Cross-platform path normalization (camino)
 ├── text.rs            # Text processing (synonym expansion, morphological variants)

package/bin/run.js

@@ -60,41 +60,10 @@ if (!env.METRICS_ENABLED) {
     env.METRICS_ENABLED = 'false';
 }
 
-// 5. Set persistence paths to be inside the project (BASE_DIR/.cimcp) 
-// if not explicitly overridden. This keeps indexes local to the project.
-const cimcpDir = path.join(env.BASE_DIR, '.cimcp');
-
-// Ensure .cimcp directory exists
-if (!fs.existsSync(cimcpDir)) {
-    try {
-        fs.mkdirSync(cimcpDir, { recursive: true });
-    } catch (e) {
-        console.error(`Failed to create .cimcp directory at ${cimcpDir}:`, e.message);
-        // Continue, the server might handle it or fail later
-    }
-}
-
-if (!env.DB_PATH) env.DB_PATH = path.join(cimcpDir, 'code-intelligence.db');
-if (!env.VECTOR_DB_PATH) env.VECTOR_DB_PATH = path.join(cimcpDir, 'vectors');
-if (!env.TANTIVY_INDEX_PATH) env.TANTIVY_INDEX_PATH = path.join(cimcpDir, 'tantivy-index');
-
-// Also set model dir - use GLOBAL cache to avoid downloading models for every project
-// Models are shared across projects, but indexes remain local
-if (!env.EMBEDDINGS_MODEL_DIR) {
-    // macOS: ~/Library/Application Support/cimcp/embeddings-cache
-    env.EMBEDDINGS_MODEL_DIR = path.join(os.homedir(), 'Library', 'Application Support', 'cimcp', 'embeddings-cache');
-
-    // Ensure global model cache directory exists
-    if (!fs.existsSync(env.EMBEDDINGS_MODEL_DIR)) {
-        try {
-            fs.mkdirSync(env.EMBEDDINGS_MODEL_DIR, { recursive: true });
-        } catch (e) {
-            console.error(`Failed to create global embeddings cache at ${env.EMBEDDINGS_MODEL_DIR}:`, e.message);
-            console.warn('Falling back to local project cache for this session');
-            env.EMBEDDINGS_MODEL_DIR = path.join(cimcpDir, 'embeddings-model');
-        }
-    }
-}
+// Per-repo data paths are auto-derived by the Rust binary from BASE_DIR.
+// Indexes are stored under ~/.code-intelligence/repos/<hash>/.
+// Models are shared across projects under ~/.code-intelligence/models/.
+// No need to set DB_PATH, VECTOR_DB_PATH, or TANTIVY_INDEX_PATH here.
 
 // Spawn the process
 const child = spawn(BINARY_PATH, process.argv.slice(2), {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@iceinvein/code-intelligence-mcp",
-  "version": "1.1.3",
+  "version": "1.2.0",
   "description": "Code Intelligence MCP Server - Smart context for your LLM coding agent",
   "bin": {
     "code-intelligence-mcp": "bin/run.js",