Update inference.py

inference.py

@@ -2,12 +2,14 @@ import os
 import torch
 import torch.nn.functional as F
 from transformers import AutoTokenizer
-from evo_model import EvoTransformerV22
+from evo_model import build_model_from_config
+from evo_architecture import mutate_genome, default_config, log_genome
 from search_utils import web_search
 import openai
 import time
 import psutil
 import platform
+import csv
 
 openai.api_key = os.getenv("OPENAI_API_KEY")
 
@@ -15,14 +17,16 @@ MODEL_PATH = "evo_hellaswag.pt"
 tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
 model = None
 last_mod_time = 0
+current_config = default_config()
 
-# 🔁 Load Evo model with auto-reload
-def load_model():
-    global model, last_mod_time
+
+# 🔁 Load Evo model with auto-reload and dynamic config
+def load_model(force_reload=False):
+    global model, last_mod_time, current_config
     try:
         current_mod_time = os.path.getmtime(MODEL_PATH)
-        if model is None or current_mod_time > last_mod_time:
-            model = EvoTransformerV22()
+        if model is None or force_reload or current_mod_time > last_mod_time:
+            model = build_model_from_config(current_config)
             model.load_state_dict(torch.load(MODEL_PATH, map_location="cpu"))
             model.eval()
             last_mod_time = current_mod_time
@@ -32,6 +36,7 @@ def load_model():
         model = None
     return model
 
+
 # 🔮 Evo inference core logic
 def evo_infer(query, options, user_context=""):
     model = load_model()
@@ -49,8 +54,8 @@ def evo_infer(query, options, user_context=""):
         context_str = "\n".join(search_results + ([user_context] if user_context else []))
 
     input_pairs = [f"{query} [SEP] {opt} [CTX] {context_str}" for opt in options]
-
     scores = []
+
     for pair in input_pairs:
         encoded = tokenizer(pair, return_tensors="pt", padding="max_length", truncation=True, max_length=128)
         with torch.no_grad():
@@ -66,6 +71,7 @@ def evo_infer(query, options, user_context=""):
         context_str
     )
 
+
 # 🤖 GPT fallback (for comparison)
 def get_gpt_response(query, user_context=""):
     try:
@@ -79,6 +85,7 @@ def get_gpt_response(query, user_context=""):
     except Exception as e:
         return f"⚠️ GPT error:\n{str(e)}"
 
+
 # 🧠 Live Evo prediction logic
 def evo_chat_predict(history, query, options):
     try:
@@ -99,16 +106,11 @@ def evo_chat_predict(history, query, options):
         "context_used": evo_ctx
     }
 
+
 # 📊 Evo model config metadata
 def get_model_config():
-    return {
-        "num_layers": 6,
-        "num_heads": 8,
-        "ffn_dim": 1024,
-        "memory_enabled": True,
-        "phase": "v2.2",
-        "accuracy": "~64.5%"
-    }
+    return current_config
+
 
 # 🖥️ Runtime stats
 def get_system_stats():
@@ -125,22 +127,41 @@ def get_system_stats():
         "platform": platform.platform()
     }
 
-# 🔁 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()
-    if model is None:
-        return "❌ Evo model not available."
 
+# 🔁 Retrain from feedback_log.csv and evolve architecture
+def retrain_from_feedback_csv():
+    global current_config, model
+    path = "feedback_log.csv"
+    if not os.path.exists(path):
+        return "⚠️ No feedback_log.csv found."
+
+    feedback_data = []
+    with open(path, newline='', encoding="utf-8") as f:
+        reader = csv.DictReader(f)
+        for row in reader:
+            q = row["question"]
+            o1 = row["option1"]
+            o2 = row["option2"]
+            ctx = row["context"]
+            vote = row.get("user_preference", "").lower()
+            correct = row.get("evo_was_correct", "").lower()
+            if vote == "evo" or correct == "yes":
+                feedback_data.append((q, o1, o2, o2))  # Evo was correct
+            elif vote == "gpt":
+                feedback_data.append((q, o1, o2, o1))  # Evo was wrong
+
+    if not feedback_data:
+        return "⚠️ No usable feedback data."
+
+    # Mutate and retrain new model
+    current_config = mutate_genome(current_config)
+    model = build_model_from_config(current_config)
     model.train()
     optimizer = torch.optim.Adam(model.parameters(), lr=1e-4)
 
-    for row in feedback_log:
-        question, opt1, opt2, answer, *_ = row
-        label = torch.tensor([1.0 if answer.strip() == opt2.strip() else 0.0])  # opt2 = class 1
-
+    for row in feedback_data:
+        question, opt1, opt2, answer = row
+        label = torch.tensor([1.0 if answer.strip() == opt2.strip() else 0.0])
         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)
 
@@ -151,4 +172,5 @@ def retrain_from_feedback(feedback_log):
         optimizer.zero_grad()
 
     torch.save(model.state_dict(), MODEL_PATH)
-    return "✅ Evo retrained and reloaded from memory."
+    log_genome(current_config)
+    return "✅ Evo mutated, retrained, and saved."