Update inference.py

inference.py

@@ -9,10 +9,8 @@ import time
 import psutil
 import platform
 
-# 🔐 Load OpenAI API Key
 openai.api_key = os.getenv("OPENAI_API_KEY")
 
-# 📦 Constants
 MODEL_PATH = "evo_hellaswag.pt"
 tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
 model = None
@@ -40,7 +38,6 @@ def evo_infer(query, options, user_context=""):
     if model is None:
         return "Model Error", 0.0, "Model not available", ""
 
-    # ✅ Smart context logic: avoid search for numeric/logical queries
     def is_fact_or_math(q):
         q_lower = q.lower()
         return any(char.isdigit() for char in q_lower) or any(op in q_lower for op in ["+", "-", "*", "/", "=", "what is", "solve", "calculate"])
@@ -69,8 +66,7 @@ def evo_infer(query, options, user_context=""):
         context_str
     )
 
-
-# 💬 GPT fallback (used for comparison only)
+# 🤖 GPT fallback (for comparison)
 def get_gpt_response(query, user_context=""):
     try:
         context_block = f"\n\nContext:\n{user_context}" if user_context else ""
@@ -83,10 +79,9 @@ def get_gpt_response(query, user_context=""):
     except Exception as e:
         return f"⚠️ GPT error:\n{str(e)}"
 
-# 🤖 Evo live chat prediction
+# 🧠 Live Evo prediction logic
 def evo_chat_predict(history, query, options):
     try:
-        # Support list or DataFrame
         if isinstance(history, list):
             context = "\n".join(history[-6:])
         elif hasattr(history, "empty") and not history.empty:
@@ -104,7 +99,7 @@ def evo_chat_predict(history, query, options):
         "context_used": evo_ctx
     }
 
-# 📊 Evo architecture stats
+# 📊 Evo model config metadata
 def get_model_config():
     return {
         "num_layers": 6,
@@ -115,7 +110,7 @@ def get_model_config():
         "accuracy": "~64.5%"
     }
 
-# 🖥️ System runtime stats
+# 🖥️ Runtime stats
 def get_system_stats():
     gpu_info = torch.cuda.get_device_properties(0) if torch.cuda.is_available() else None
     memory = psutil.virtual_memory()
@@ -130,9 +125,9 @@ def get_system_stats():
         "platform": platform.platform()
     }
 
-# 🧪 Fine-tune Evo from feedback data (CSV or in-memory list)
-def retrain_from_feedback(feedback_data):
-    if not feedback_data:
+# 🔁 Retrain from in-memory feedback_log
+def retrain_from_feedback(feedback_log):
+    if not feedback_log:
         return "⚠️ No feedback data to retrain from."
 
     model = load_model()
@@ -142,11 +137,10 @@ def retrain_from_feedback(feedback_data):
     model.train()
     optimizer = torch.optim.Adam(model.parameters(), lr=1e-4)
 
-    for row in feedback_data:
+    for row in feedback_log:
         question, opt1, opt2, answer, *_ = row
-        label = torch.tensor([1.0 if answer.strip() == opt2.strip() else 0.0])  # opt2 is class 1
+        label = torch.tensor([1.0 if answer.strip() == opt2.strip() else 0.0])  # opt2 = class 1
 
-        # Build input pair
         input_text = f"{question} [SEP] {opt2 if label.item() == 1 else opt1}"
         encoded = tokenizer(input_text, return_tensors="pt", padding="max_length", truncation=True, max_length=128)
 
@@ -157,5 +151,4 @@ def retrain_from_feedback(feedback_data):
         optimizer.zero_grad()
 
     torch.save(model.state_dict(), MODEL_PATH)
-    return "✅ Evo retrained from feedback."
-
+    return "✅ Evo retrained and reloaded from memory."