inference.py

import os
import torch
from evo_model import EvoTransformer
from transformers import AutoTokenizer
from rag_utils import extract_text_from_file
from search_utils import web_search

# Load Evo model
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
model = EvoTransformer()
model.load_state_dict(torch.load("evo_hellaswag.pt", map_location="cpu"))
model.eval()

def get_evo_response(query, context=None, enable_search=True):
    search_snippets = ""
    if enable_search:
        snippets = web_search(query)
        if snippets:
            search_snippets = "\n".join(snippets)

    full_context = f"{context or ''}\n\n{search_snippets}".strip()
    input_1 = f"{query} Option 1"
    input_2 = f"{query} Option 2"

    inputs = tokenizer([input_1, input_2], padding=True, truncation=True, return_tensors="pt")
    with torch.no_grad():
        logits = model(inputs["input_ids"]).squeeze(-1)
        probs = torch.softmax(logits, dim=0)
        best_idx = torch.argmax(probs).item()

    suggestion = f"Option {best_idx + 1}"
    reasoning = (
        f"Evo suggests: **{suggestion}** (Confidence: {probs[best_idx]:.2f})\n\n"
        f"Context used:\n{full_context}"
    )
    return suggestion, reasoning

def get_gpt_response(query, context=None):
    import openai
    openai.api_key = os.getenv("OPENAI_API_KEY", "")
    context = context or "None"

    try:
        response = openai.ChatCompletion.create(
            model="gpt-3.5-turbo",
            messages=[
                {"role": "system", "content": "You are a helpful expert advisor."},
                {"role": "user", "content": f"Context: {context}\n\nQuestion: {query}"}
            ],
            max_tokens=250
        )
        return response["choices"][0]["message"]["content"].strip()
    except Exception as e:
        return f"⚠️ GPT error: {str(e)}"