updated RAG_4

fiscal.py

@@ -26,24 +26,40 @@ except ImportError:
 # ---------- Embeddings ----------
 _embed_lock = threading.Lock()
 
-embeddings = HuggingFaceEmbeddings(
+_raw_embeddings = HuggingFaceEmbeddings(
     model_name="nomic-ai/nomic-embed-text-v1",
     model_kwargs={"trust_remote_code": True},
 )
 
-_orig_embed_query = embeddings.embed_query
-_orig_embed_documents = embeddings.embed_documents
 
-def _locked_embed_query(text):
-    with _embed_lock:
-        return _orig_embed_query(text)
+class LockedEmbeddings:
+    """Serializes access to the underlying embedding model.
+
+    The Nomic model caches rotary cos/sin buffers on itself and resizes them
+    in place, which is not safe when two requests embed concurrently.
+    """
+
+    def __init__(self, inner):
+        self._inner = inner
+
+    def embed_query(self, text):
+        with _embed_lock:
+            return self._inner.embed_query(text)
+
+    def embed_documents(self, texts):
+        with _embed_lock:
+            return self._inner.embed_documents(texts)
+
+    async def aembed_query(self, text):
+        with _embed_lock:
+            return self._inner.embed_query(text)
+
+    async def aembed_documents(self, texts):
+        with _embed_lock:
+            return self._inner.embed_documents(texts)
 
-def _locked_embed_documents(texts):
-    with _embed_lock:
-        return _orig_embed_documents(texts)
 
-embeddings.embed_query = _locked_embed_query
-embeddings.embed_documents = _locked_embed_documents
+embeddings = LockedEmbeddings(_raw_embeddings)
 
 # ---------- Vector store ----------
 vectorstore = Chroma(