main.py

# from fastapi import FastAPI
# from pydantic import BaseModel
# from fastapi.openapi.utils import get_openapi
# from transformers import AutoTokenizer, AutoModelForCausalLM
# import torch

# app = FastAPI(
#     title="Harshal AI Backend",
#     version="1.0.0",
# )

# MODEL_NAME = "Qwen/Qwen2.5-0.5B"

# tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
# model = AutoModelForCausalLM.from_pretrained(
#     MODEL_NAME,
#     torch_dtype=torch.float32,
#     device_map="cpu",
# )

# class ChatMessage(BaseModel):
#     messages: list

# @app.get("/")
# def home():
#     return {"message": "Harshal AI backend running with Qwen 0.5B!"}

# @app.post("/chat")
# def chat(body: ChatMessage):
#     user_msg = body.messages[-1]["content"]
#     prompt = f"User: {user_msg}\nAssistant:"

#     inputs = tokenizer(prompt, return_tensors="pt")
#     outputs = model.generate(
#         **inputs,
#         max_new_tokens=120,
#         pad_token_id=tokenizer.eos_token_id,
#         temperature=0.4,
#     )

#     text = tokenizer.decode(outputs[0], skip_special_tokens=True)
#     reply = text.split("Assistant:")[-1].strip()
#     return {"reply": reply}

# @app.get("/openapi.json")
# def openapi_json():
#     return get_openapi(
#         title="Harshal AI Backend",
#         version="1.0.0",
#         routes=app.routes
#     )


from fastapi import FastAPI
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel
from typing import List
from transformers import AutoTokenizer, AutoModelForCausalLM
from sentence_transformers import SentenceTransformer
from pypdf import PdfReader
import torch
import os

app = FastAPI(
    title="Harshal AI Backend",
    version="1.0.0"
)

# CORS so Next.js can call backend
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_methods=["*"],
    allow_headers=["*"],
)

# ======================================
# 1) LOAD MAIN MODEL (Qwen2.5 1.5B)
# ======================================
MODEL_NAME = "Qwen/Qwen2.5-1.5B-Instruct"

tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
llm = AutoModelForCausalLM.from_pretrained(
    MODEL_NAME,
    torch_dtype=torch.float32,
    device_map="cpu"
)
llm.eval()

# ======================================
# 2) LOAD EMBEDDING MODEL + RAG
# ======================================
EMBED_MODEL = "sentence-transformers/all-MiniLM-L6-v2"
embedder = SentenceTransformer(EMBED_MODEL)

RESUME = "resume.pdf"
resume_rag = None


def chunk_text(text, max_chars=450, overlap=80):
    text = " ".join(text.split())
    chunks, start = [], 0

    while start < len(text):
        end = start + max_chars
        chunks.append(text[start:end])
        start = end - overlap

    return chunks


def build_rag():
    global resume_rag

    if not os.path.exists(RESUME):
        print("⚠ resume.pdf NOT FOUND — RAG disabled")
        return

    reader = PdfReader(RESUME)
    text = ""

    for p in reader.pages:
        text += p.extract_text() or ""

    chunks = chunk_text(text)

    embs = embedder.encode(chunks, convert_to_tensor=True, normalize_embeddings=True)

    resume_rag = {"chunks": chunks, "embs": embs}
    print("✅ RAG Ready with", len(chunks), "chunks")


def get_rag_context(query):
    if resume_rag is None:
        return ""

    q = embedder.encode([query], convert_to_tensor=True, normalize_embeddings=True)[0]
    sims = torch.nn.functional.cosine_similarity(q.unsqueeze(0), resume_rag["embs"])
    top = torch.topk(sims, k=3)

    return "\n\n".join(resume_rag["chunks"][i] for i in top.indices)


build_rag()

# ======================================
# 3) SWAGGER DATA MODELS
# ======================================
class Msg(BaseModel):
    role: str
    content: str

class ChatReq(BaseModel):
    messages: List[Msg]


# ======================================
# 4) CHAT ROUTE (Human-like Harshal)
# ======================================
@app.post("/chat")
def chat(req: ChatReq):

    user_msg = req.messages[-1].content.strip()

    rag = get_rag_context(user_msg)

    persona = f"""
You are Harshal Sonawane, a real human software engineer from Pune.

Rules:
- Do NOT say you are an AI or model.
- Answer in a human, natural tone.
- Keep replies short (2–4 sentences).
- Use resume facts when relevant.
- If unknown, answer honestly.

Resume context:
{rag}
""".strip()

    messages = [
        {"role": "system", "content": persona},
        {"role": "user", "content": user_msg}
    ]

    ids = tokenizer.apply_chat_template(
        messages,
        return_tensors="pt",
        add_generation_prompt=True
    ).to(llm.device)

    out = llm.generate(
        ids,
        max_new_tokens=150,
        temperature=0.5,
        top_p=0.9,
        repetition_penalty=1.05,
        do_sample=True
    )

    gen = out[0][ids.shape[-1]:]
    reply = tokenizer.decode(gen, skip_special_tokens=True).strip()

    return {"reply": reply}


@app.get("/")
def health():
    return {"status": "Harshal AI backend is running with Qwen 1.5B + RAG 🎯"}