Update inference.py

inference.py

@@ -1,6 +1,7 @@
 # inference.py
 
 import re
+import threading
 import pandas as pd
 import torch
 import faiss
@@ -60,8 +61,14 @@ faiss.normalize_L2(row_embs)
 row_idx = faiss.IndexFlatIP(row_embs.shape[1])
 row_idx.add(row_embs)
 
+# ─────────────────────────────────────────────────────────────────────────────
+#  ⚙️ Thread-safety için kilitler
+faiss_lock = threading.Lock()
+model_lock = threading.Lock()
+
 # ─────────────────────────────────────────────────────────────────────────────
 # 5️⃣ generate_answer: app.py’in çağıracağı fonksiyon
+@torch.inference_mode()
 def generate_answer(user_question: str) -> str:
     # (1) normalize “makine” → “RTF makinesi”
     q_norm = re.sub(r"\bmakine\b", "RTF makinesi", user_question, flags=re.IGNORECASE)
@@ -70,7 +77,8 @@ def generate_answer(user_question: str) -> str:
     if not re.search(r"\b(makine|titreşim|alarm|rtf)\b", q_norm, flags=re.IGNORECASE):
         prompt = SYSTEM_PREFIX + "\n" + f"Soru: {q_norm}\nCevap:"
         inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
-        out    = model.generate(**inputs, max_new_tokens=1000)
+        with model_lock:
+            out = model.generate(**inputs, max_new_tokens=1000)
         return tokenizer.decode(out[0], skip_special_tokens=True)
 
     # (3a) veri‐ilgili ise önce static QA
@@ -84,11 +92,13 @@ def generate_answer(user_question: str) -> str:
         date=date
     )
 
-    # (3b) fallback
+    # (3b) fallback: static QA başarısızsa dynamic RAG + LLM
     if any(tok in ans for tok in ["Cevap bulunamadı", "Lütfen sorunuzda", "Tam olarak anlayamadım"]):
+        # FAISS üzerinden ilgili satırları al
         ue = embedder_q.encode([q_norm], convert_to_numpy=True)
         faiss.normalize_L2(ue)
-        D_rows, I_rows = row_idx.search(ue, 5)
+        with faiss_lock:
+            D_rows, I_rows = row_idx.search(ue, 5)
         context = "\n".join(row_texts[i] for i in I_rows[0])
 
         prompt = (
@@ -99,7 +109,8 @@ def generate_answer(user_question: str) -> str:
             "Bu verilere dayanarak cevap verin:"
         )
         inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
-        out    = model.generate(**inputs, max_new_tokens=1000)
+        with model_lock:
+            out = model.generate(**inputs, max_new_tokens=1000)
         return tokenizer.decode(out[0], skip_special_tokens=True)
 
     # (3c) static QA cevabı