Add initial GraphRAG implementation

.claude/settings.local.json

@@ -1,6 +1,10 @@
 {
-  "enabledMcpjsonServers": ["codeembed"],
+  "enabledMcpjsonServers": [
+    "codeembed"
+  ],
   "permissions": {
-    "allow": ["mcp__codeembed__search"]
+    "allow": [
+      "mcp__codeembed__search"
+    ]
   }
 }

.mcp.json

@@ -1,8 +1,10 @@
 {
   "mcpServers": {
     "codeembed": {
-      "command": "uv",
-      "args": ["run", "codeembed", "serve"]
+      "command": "codeembed",
+      "args": [
+        "serve"
+      ]
     }
   }
 }

.vscode/mcp.json

@@ -1,8 +1,10 @@
 {
   "servers": {
     "codeembed": {
-      "command": "uv",
-      "args": ["run", "codeembed", "serve"]
+      "command": "codeembed",
+      "args": [
+        "serve"
+      ]
     }
   }
 }

README.md

@@ -1,6 +1,6 @@
 # CodeEmbed
 
-Embeds your codebase into a local vector database and exposes it as an MCP tool, giving AI assistants like Claude Code fast semantic search over your code.
+Embeds your codebase into a local vector and graph database and exposes it as an MCP tool, giving AI assistants like Claude Code fast semantic search over your code using Graph RAG.
 
 Particularly useful for questions like:
 
@@ -12,7 +12,7 @@ For other questions, the agent will fall back to normal lookups.
 CodeEmbed can improve lookup speed and accuracy, especially for finding existing implementations before writing new ones.
 Note that the biggest bottleneck in coding agents is LLM thinking and token generation — solid prompts and follow-up questions still matter.
 
-Uses [ChromaDB](https://github.com/chroma-core/chroma) for local vector storage and either [Ollama](https://github.com/ollama/ollama) or OpenAI (including OpenAI models via Azure AI Foundry) for LLM analysis.
+Uses [ChromaDB](https://github.com/chroma-core/chroma) for vector storage, SQLite for graph storage, and either [Ollama](https://github.com/ollama/ollama) or OpenAI (including OpenAI models via Azure AI Foundry) for LLM analysis.
 
 ## Prerequisites
 

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "codeembed"
-version = "0.1.1"
+version = "0.2.0b3"
 description = "Embeds your codebase and makes it available for quick LLM lookups via MCP."
 readme = "README.md"
 requires-python = ">=3.11"

src/codeembed/bootstrap/services.py

@@ -10,6 +10,7 @@
 from codeembed.doc_embedder.doc_embedder import DocEmbedder
 from codeembed.doc_provider.local_doc_provider import LocalDocProvider
 from codeembed.doc_search_service.doc_search_service import DocSearchService
+from codeembed.graph_db.sqlite_adapter import SqliteGraphDb
 from codeembed.llm.base import LLMServiceBase
 from codeembed.llm.ollama_adapter import OllamaLLMService
 from codeembed.vector_db.chromadb_adapter import ChromaDbAdapter
@@ -23,10 +24,21 @@
 _DEFAULT_SLEEP_INTERVAL = 60
 
 
+@lru_cache(maxsize=1)
+def get_vector_db() -> ChromaDbAdapter:
+    return ChromaDbAdapter(collection_name="codebase")
+
+
+@lru_cache(maxsize=1)
+def get_graph_db() -> SqliteGraphDb:
+    return SqliteGraphDb(db_path=".codeembed/graph.db")
+
+
 @lru_cache(maxsize=1)
 def get_search_service() -> DocSearchService:
-    vector_db = ChromaDbAdapter(collection_name="codebase")
-    search_service = DocSearchService(vector_db)
+    vector_db = get_vector_db()
+    graph_db = get_graph_db()
+    search_service = DocSearchService(vector_db, graph_db)
     return search_service
 
 
@@ -195,10 +207,11 @@ def get_embedder_service() -> DocEmbedder:
         base_path=".",
         supported_file_extensions=_SUPPORTED_FILE_EXTENSIONS,
     )
-    vector_db = ChromaDbAdapter(collection_name="codebase")
+    vector_db = get_vector_db()
     llm_service = get_llm_service()
+    graph_db = get_graph_db()
     embedder = DocEmbedder(
-        doc_provider, vector_db, llm_service, llm_model=config.llm_model, debounce_seconds=config.debounce
+        doc_provider, vector_db, graph_db, llm_service, llm_model=config.llm_model, debounce_seconds=config.debounce
     )
     return embedder
 

src/codeembed/cli.py

@@ -418,6 +418,18 @@ def serve():
     mcp.run(transport="stdio")
 
 
+@app.command()
+def search(
+    query: str = typer.Argument(..., help="Natural-language search query"),
+    top_n: int = typer.Option(10, "--top-n", "-n", help="Number of results to return"),
+):
+    """Search the embedded codebase using semantic similarity."""
+    from codeembed.bootstrap.services import get_search_service
+
+    result = get_search_service().search(query, top_n)
+    typer.echo(result)
+
+
 @app.command()
 def embed():
     """Embed codebase into the vector database."""

src/codeembed/delta_computer/delta_computer.py

@@ -15,7 +15,7 @@ def __init__(self, doc_provider: DocProviderBase, vector_db: VectorDbBase, debou
         self._vector_db = vector_db
         self._debounce_seconds = debounce_seconds
 
-    def compute_deltas(self) -> Tuple[Set[UUID], Set[str]]:
+    def compute_deltas(self) -> Tuple[Set[UUID], Set[str], Set[str]]:
         """
         Returns chunk IDs to delete and file paths to process.
 
@@ -26,6 +26,7 @@ def compute_deltas(self) -> Tuple[Set[UUID], Set[str]]:
 
         file_path_to_chunk_ids: Dict[str, List[UUID]] = {}
         chunk_ids_to_delete: Set[UUID] = set()
+        file_paths_to_delete: Set[str] = set()
 
         # Collect modified_at stored in our database.
         old_modified_at: Dict[str, datetime] = {}
@@ -71,5 +72,6 @@ def compute_deltas(self) -> Tuple[Set[UUID], Set[str]]:
             if file_path not in current:
                 for chunk_id in file_path_to_chunk_ids.get(file_path, []):
                     chunk_ids_to_delete.add(chunk_id)
+                file_paths_to_delete.add(file_path)
 
-        return chunk_ids_to_delete, file_paths_to_update
+        return chunk_ids_to_delete, file_paths_to_update, file_paths_to_delete