Mirrors/talemate
1
0
Fork 0
mirror of https://github.com/vegu-ai/talemate.git synced 2026-09-02 03:59:12 +02:00
Code Issues Packages Projects Releases Wiki Activity

fix: optimize model eviction logic in ChromaDBMemoryAgent

Browse Source
This commit is contained in:
vegu-ai-tools
2026-04-14 23:57:39 +03:00
parent 4e82220de8
commit 5f85cc532d
1 changed files with 8 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
src/talemate/agents/memory/__init__.py
View File

@@ -968,11 +968,17 @@ class ChromaDBMemoryAgent(MemoryAgent):
)
try:
# Evict cached model so device changes take effect without restart.
# ChromaDB caches models in a class-level dict keyed only by model name,
# ignoring device — so a device switch would silently reuse the old model.
# Evict only when the cached model's device differs from the requested one,
# otherwise reuse it (avoids an expensive reload on every set_db call).
ST = embedding_functions.SentenceTransformerEmbeddingFunction
if model_name in ST.models:
cached = ST.models.get(model_name)
if (
cached is not None
and getattr(cached, "device", None) is not None
and cached.device.type != device
):
self._release_embedding_model(ST.models.pop(model_name))
ef = ST(