fix: share a single BM25 model between indexing and search

server/main.py

@@ -11,14 +11,16 @@
 
 from server.config import settings
 from server.embeddings import close_embedding_provider, get_embedding_provider
-from server.embeddings.bm25 import BM25SparseProvider, close_sparse_embedding_provider
+from server.embeddings.bm25 import (
+    close_sparse_embedding_provider,
+    get_sparse_embedding_provider,
+)
 from server.state import (
     get_commit_store,
     get_service_registry,
     get_store,
     set_commit_store,
     set_service_registry,
-    set_sparse_provider,
     set_store,
 )
 from server.store.commit_store import CommitStore
@@ -47,8 +49,9 @@ async def lifespan(_: MCPServer) -> AsyncIterator[None]:
     await commit_store.ensure_collection()
     set_commit_store(commit_store)
 
-    sparse_provider = BM25SparseProvider()
-    set_sparse_provider(sparse_provider)
+    # Warm the single BM25 model at startup so the first query does not pay the
+    # load cost; indexing and search share this one instance.
+    get_sparse_embedding_provider()
 
     set_service_registry(ServiceRegistry())
 

server/state.py

@@ -3,14 +3,12 @@
 import asyncio
 from collections import defaultdict
 
-from server.embeddings.bm25 import BM25SparseProvider
 from server.store.commit_store import CommitStore
 from server.store.qdrant import QdrantStore
 from server.store.service_registry import ServiceRegistry
 
 _store: QdrantStore | None = None
 _commit_store: CommitStore | None = None
-_sparse_provider: BM25SparseProvider | None = None
 _service_registry: ServiceRegistry | None = None
 _reindex_locks: defaultdict[str, asyncio.Lock] = defaultdict(asyncio.Lock)
 
@@ -37,17 +35,6 @@ def set_commit_store(store: CommitStore) -> None:
     _commit_store = store
 
 
-def get_sparse_provider() -> BM25SparseProvider:
-    if _sparse_provider is None:
-        raise RuntimeError("Sparse embedding provider not initialized")
-    return _sparse_provider
-
-
-def set_sparse_provider(provider: BM25SparseProvider) -> None:
-    global _sparse_provider
-    _sparse_provider = provider
-
-
 def get_service_registry() -> ServiceRegistry:
     if _service_registry is None:
         raise RuntimeError("Service registry not initialized")

server/tools/search.py

@@ -7,8 +7,9 @@
 
 from server.config import settings
 from server.embeddings import get_embedding_provider
+from server.embeddings.bm25 import get_sparse_embedding_provider
 from server.indexer.github_source import fetch_file_content
-from server.state import get_service_registry, get_sparse_provider, get_store
+from server.state import get_service_registry, get_store
 from server.store.service_registry import load_effective_services
 from server.tools.file_cache import BlobContentCache
 
@@ -44,7 +45,7 @@ async def search_code(
             limit: Maximum number of results (default 10)
         """
         embedder = get_embedding_provider()
-        sparse_embedder = get_sparse_provider()
+        sparse_embedder = get_sparse_embedding_provider()
         store = get_store()
 
         dense_vector = await embedder.embed_query(query)
@@ -147,7 +148,7 @@ async def find_usages(
             limit: Maximum number of results (default 10)
         """
         embedder = get_embedding_provider()
-        sparse_embedder = get_sparse_provider()
+        sparse_embedder = get_sparse_embedding_provider()
         store = get_store()
 
         query = f"code that uses or references {symbol_name}"

tests/embeddings/test_bm25_singleton.py

@@ -0,0 +1,49 @@
+from __future__ import annotations
+
+import pytest
+
+import server.embeddings.bm25 as bm25_module
+import server.indexer.pipeline as pipeline_module
+import server.tools.search as search_module
+from server.embeddings.bm25 import (
+    close_sparse_embedding_provider,
+    get_sparse_embedding_provider,
+)
+
+
+class _StubBm25:
+    def __init__(self, model_name: str) -> None:
+        self.model_name = model_name
+
+
+@pytest.fixture(autouse=True)
+def stub_model(monkeypatch):
+    """Keep the tests off the real fastembed model download."""
+    monkeypatch.setattr(bm25_module, "Bm25", _StubBm25)
+    monkeypatch.setattr(bm25_module, "_provider", None)
+
+
+def test_get_sparse_embedding_provider_returns_singleton() -> None:
+    assert get_sparse_embedding_provider() is get_sparse_embedding_provider()
+
+
+def test_indexing_and_search_share_one_instance() -> None:
+    """Guards against reintroducing a second BM25 holder (issue #70): the
+    pipeline and the search tools must resolve to the same model."""
+    assert (
+        pipeline_module.get_sparse_embedding_provider
+        is search_module.get_sparse_embedding_provider
+    )
+    assert (
+        pipeline_module.get_sparse_embedding_provider()
+        is search_module.get_sparse_embedding_provider()
+    )
+
+
+async def test_close_releases_the_only_instance() -> None:
+    provider = get_sparse_embedding_provider()
+
+    await close_sparse_embedding_provider()
+
+    assert bm25_module._provider is None
+    assert get_sparse_embedding_provider() is not provider

tests/tools/test_search.py

@@ -20,7 +20,7 @@ async def test_search_code_reports_no_results() -> None:
 
     with (
         patch("server.tools.search.get_embedding_provider") as mock_embedder,
-        patch("server.tools.search.get_sparse_provider") as mock_sparse,
+        patch("server.tools.search.get_sparse_embedding_provider") as mock_sparse,
         patch("server.tools.search.get_store", return_value=store),
     ):
         mock_embedder.return_value.embed_query = AsyncMock(return_value=[0.1])
@@ -53,7 +53,7 @@ async def test_search_code_formats_hits() -> None:
 
     with (
         patch("server.tools.search.get_embedding_provider") as mock_embedder,
-        patch("server.tools.search.get_sparse_provider") as mock_sparse,
+        patch("server.tools.search.get_sparse_embedding_provider") as mock_sparse,
         patch("server.tools.search.get_store", return_value=store),
     ):
         mock_embedder.return_value.embed_query = AsyncMock(return_value=[0.1])
@@ -74,7 +74,7 @@ async def test_search_code_passes_chunk_tier_to_store() -> None:
 
     with (
         patch("server.tools.search.get_embedding_provider") as mock_embedder,
-        patch("server.tools.search.get_sparse_provider") as mock_sparse,
+        patch("server.tools.search.get_sparse_embedding_provider") as mock_sparse,
         patch("server.tools.search.get_store", return_value=store),
     ):
         mock_embedder.return_value.embed_query = AsyncMock(return_value=[0.1])
@@ -169,7 +169,7 @@ async def test_find_usages_excludes_hits_matching_symbol_itself() -> None:
 
     with (
         patch("server.tools.search.get_embedding_provider") as mock_embedder,
-        patch("server.tools.search.get_sparse_provider") as mock_sparse,
+        patch("server.tools.search.get_sparse_embedding_provider") as mock_sparse,
         patch("server.tools.search.get_store", return_value=store),
     ):
         mock_embedder.return_value.embed_query = AsyncMock(return_value=[0.1])
@@ -187,7 +187,7 @@ async def test_find_usages_over_fetches_to_absorb_self_match_filtering() -> None
 
     with (
         patch("server.tools.search.get_embedding_provider") as mock_embedder,
-        patch("server.tools.search.get_sparse_provider") as mock_sparse,
+        patch("server.tools.search.get_sparse_embedding_provider") as mock_sparse,
         patch("server.tools.search.get_store", return_value=store),
     ):
         mock_embedder.return_value.embed_query = AsyncMock(return_value=[0.1])
@@ -227,7 +227,7 @@ async def test_find_usages_still_returns_limit_results_when_definition_ranks_fir
 
     with (
         patch("server.tools.search.get_embedding_provider") as mock_embedder,
-        patch("server.tools.search.get_sparse_provider") as mock_sparse,
+        patch("server.tools.search.get_sparse_embedding_provider") as mock_sparse,
         patch("server.tools.search.get_store", return_value=store),
     ):
         mock_embedder.return_value.embed_query = AsyncMock(return_value=[0.1])
@@ -257,7 +257,7 @@ async def test_find_usages_snippet_windows_around_match() -> None:
 
     with (
         patch("server.tools.search.get_embedding_provider") as mock_embedder,
-        patch("server.tools.search.get_sparse_provider") as mock_sparse,
+        patch("server.tools.search.get_sparse_embedding_provider") as mock_sparse,
         patch("server.tools.search.get_store", return_value=store),
     ):
         mock_embedder.return_value.embed_query = AsyncMock(return_value=[0.1])