fix(mcp_server): pass embedding_function= on collection reopen (#1299)

`mcp_server._get_collection` bypassed `ChromaBackend.get_collection`
and called `client.get_collection` / `client.create_collection` without
`embedding_function=`. ChromaDB 1.x does not persist the EF identity
with the collection, so the MCP server's reopen silently bound
chromadb's built-in `DefaultEmbeddingFunction` while the miner / Stop
hook ingest path bound `mempalace.embedding.get_embedding_function()`.

On bleeding-edge interpreters (python 3.14 + chromadb 1.5.x on Apple
Silicon, per #1299) the default EF's lazy ONNX provider selection could
SIGSEGV the host process on first `col.add()`, killing the MCP stdio
server and leaving every subsequent tool call returning
`Connection closed` until Claude Code was relaunched. Reads worked
because `col.get(ids=...)` and metadata fetches don't invoke the EF;
the auto-ingest path worked because mining routes through the backend
abstraction. Diary writes were the consistent failure surface.

Resolve the EF up front (matching `ChromaBackend._resolve_embedding_function`)
and pass it into both reopen branches. Falls back to the chromadb default
only if `mempalace.embedding.get_embedding_function` itself raises.

Regression test patches the chromadb client class to capture
`embedding_function=` on every `get_collection` / `create_collection`
call from `_get_collection(create=True)` and `_get_collection()`, and
fails if any call omits it.

Follow-up to #1262 / #1289 (which fixed the metadata-mismatch SIGSEGV
path); this addresses the EF-mismatch SIGSEGV path on the same surface.
This commit is contained in:
Igor Lins e Silva
2026-05-01 19:34:38 -03:00
parent 5e540da06b
commit ac6c2b6af6
3 changed files with 77 additions and 2 deletions
+20 -2
View File
@@ -66,6 +66,7 @@ from .backends.chroma import ( # noqa: E402
_pin_hnsw_threads,
hnsw_capacity_status,
)
from .embedding import get_embedding_function # noqa: E402
from .query_sanitizer import sanitize_query # noqa: E402
from .searcher import search_memories # noqa: E402
from .palace_graph import ( # noqa: E402
@@ -278,6 +279,22 @@ def _get_collection(create=False):
global _collection_cache, _metadata_cache, _metadata_cache_time
try:
client = _get_client()
# ChromaDB 1.x does not persist the embedding function with the
# collection, so a reader/writer that omits ``embedding_function=``
# silently gets the chromadb-built-in default. On bleeding-edge
# interpreters (#1299: python 3.14 + chromadb 1.5.x on Apple Silicon)
# the default's lazy ONNX provider selection can SIGSEGV the host
# process on first ``col.add()``. The miner / Stop hook ingest path
# avoids this because it routes through ``ChromaBackend.get_collection``
# which resolves the EF via ``mempalace.embedding.get_embedding_function``.
# The MCP server bypassed that abstraction; mirror its behaviour so
# ``tool_diary_write`` / ``tool_add_drawer`` get the same EF as mining.
try:
ef = get_embedding_function()
except Exception:
logger.exception("Failed to build embedding function; using chromadb default")
ef = None
ef_kwargs = {"embedding_function": ef} if ef is not None else {}
if create:
# hnsw:num_threads=1 disables ChromaDB's multi-threaded ParallelFor
# HNSW insert path, which has a race in repairConnectionsForUpdate /
@@ -292,7 +309,7 @@ def _get_collection(create=False):
# below skips the metadata-comparison codepath for existing
# collections, mirroring the backend-layer fix from #1262.
try:
raw = client.get_collection(_config.collection_name)
raw = client.get_collection(_config.collection_name, **ef_kwargs)
except _ChromaNotFoundError:
raw = client.create_collection(
_config.collection_name,
@@ -301,13 +318,14 @@ def _get_collection(create=False):
"hnsw:num_threads": 1,
**_HNSW_BLOAT_GUARD,
},
**ef_kwargs,
)
_pin_hnsw_threads(raw)
_collection_cache = ChromaCollection(raw)
_metadata_cache = None
_metadata_cache_time = 0
elif _collection_cache is None:
raw = client.get_collection(_config.collection_name)
raw = client.get_collection(_config.collection_name, **ef_kwargs)
_pin_hnsw_threads(raw)
_collection_cache = ChromaCollection(raw)
_metadata_cache = None