use optimum ONNX MiniLM via low-level wrapper (no disk)

rag.py

@@ -103,18 +103,27 @@ def get_texts() -> List[str]:
         print(f"⚠ Dataset fetch failed: {e} – using empty corpus")
         return []
 
-## ------------------------------------------------------------------------------rtutu
-# ------------------------------------------------------------------
 # ------------------------------------------------------------------
 @lru_cache(maxsize=1)
 def get_vectorstore() -> FAISS:
     texts = get_texts()
 
     # --- FINAL: use optimum ONNX MiniLM (already on disk) -----------------
-    from langchain_community.embeddings import OptimumSentenceEmbeddings
-    embeddings = OptimumSentenceEmbeddings(
-        model_name="optimum/all-MiniLM-L6-v2"
-    )
+    import os
+    from optimum.pipelines import pipeline
+    from langchain.embeddings.base import Embeddings
+
+    class OptimumMiniLM(Embeddings):
+        def __init__(self):
+            self.pipe = pipeline("feature-extraction",
+                                 model="optimum/all-MiniLM-L6-v2",
+                                 device="cpu")
+        def embed_documents(self, texts):
+            return [self.pipe(t)[0][0] for t in texts]
+        def embed_query(self, text):
+            return self.embed_documents([text])[0]
+
+    embeddings = OptimumMiniLM()
     # ------------------------------------------------------------------------
 
     if not texts:                       # no data → empty FAISS