chat.py

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
import re
from db import get_user_expenses

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")

# GPT-2 variants do not define a pad token by default.
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token
if model.config.pad_token_id is None:
    model.config.pad_token_id = tokenizer.eos_token_id


def _clean_text(text):
    cleaned = re.sub(r"\s+", " ", text).strip()
    cleaned = re.sub(r"(.)\1{5,}", r"\1\1", cleaned)
    return cleaned


def format_expense_data(data):
    if not data:
        return "No expense data available."

    total = sum([d["amount"] for d in data])

    category_summary = {}
    for d in data:
        cat = d.get("category", "other")
        category_summary[cat] = category_summary.get(cat, 0) + d["amount"]

    text = f"Total spending: {total}\n"
    for cat, amt in category_summary.items():
        text += f"{cat}: {amt}\n"

    return text


def chat_response(query, access_token):
    user_data = get_user_expenses(access_token)

    context = format_expense_data(user_data)

    prompt = f"""
You are a smart financial advisor.
User financial data:
{context}
User question:
{query}
Instructions:
- Use the data to give advice
- Keep answers short and practical
- Do NOT generate unrelated content
- If user asks for a number, include the exact number from the provided data when possible
Answer:
"""

    input_ids = tokenizer.encode(prompt + tokenizer.eos_token, return_tensors='pt')

    with torch.no_grad():
        output = model.generate(
            input_ids,
            max_new_tokens=64,
            pad_token_id=tokenizer.eos_token_id,
            do_sample=True,
            temperature=0.3,
            top_p=0.9,
            no_repeat_ngram_size=3,
            repetition_penalty=1.2,
        )

    response = tokenizer.decode(
        output[:, input_ids.shape[-1]:][0],
        skip_special_tokens=True
    )

    response = response.split("Answer:")[-1].strip()
    response = response.split("\n")[0]
    response = _clean_text(response)

    # If model output is empty/noisy, fall back to a concise financial guidance line.
    if len(response) < 8:
        return "Focus on your top spending category and set a weekly cap to reduce non-essential expenses."

    return response