Update app.py

app.py

@@ -27,24 +27,32 @@ class EmbeddingBackend:
     def __init__(self, repo: str):
         self.repo = repo
         if repo == "BAAI/bge-small-en-v1.5":
-            # FlagEmbedding back‑end (BGE)
             self.model = FlagModel(
                 repo,
-                query_instruction_for_retrieval="Generate a representation for this sentence to retrieve related articles：",
+                query_instruction_for_retrieval="为这个句子生成表示以用于检索相关文章：",
                 use_fp16=True,
             )
-            self.encode_docs  = self.model.encode
+            self.encode_docs  = lambda docs: self.model.encode(docs, batch_size=BATCH_SIZE)
             self.encode_query = lambda q: self.model.encode_queries([q])[0]
         else:
-            # SentenceTransformer back‑ends
-            self.model = SentenceTransformer(repo, trust_remote_code=True)
+            model_kwargs = {}
+            if "Qwen3" in repo and not os.getenv("QWEN_USE_FLASH"):
+                model_kwargs["attn_implementation"] = "eager"  # lower‑mem CPU path
+            self.model = SentenceTransformer(repo, trust_remote_code=True, model_kwargs=model_kwargs)
+
+            # Custom token truncation handled externally
             if "Qwen3" in repo:
                 self.encode_query = lambda q: self.model.encode(q, prompt_name="query")
             elif "stella" in repo:
                 self.encode_query = lambda q: self.model.encode(q, prompt_name="s2p_query")
             else:
                 self.encode_query = lambda q: self.model.encode(q)
-            self.encode_docs = lambda docs: self.model.encode(docs)
+
+            self.encode_docs = lambda docs: self.model.encode(
+                docs,
+                batch_size=BATCH_SIZE,
+                normalize_embeddings=False,
+            )
 
     # Convenience wrappers that return *numpy* arrays
     def encode_corpus(self, passages: List[str]) -> np.ndarray: