main.py

from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
from transformers import AutoModelForCausalLM, AutoTokenizer
from TextToSpeech import text_to_speech
import torch
import base64

# =========================================
# ENUM MAPPINGS (Match Backend Enums)
# =========================================

SESSION_TYPES = {
    1: "technical",
    2: "softskills"
}

TRACKS = {
    19: "generalprogramming"
}

# =========================================
# LOAD MODEL ONCE (Global)
# =========================================

MODEL_PATH = "Fayza38/Question_and_Answer"

tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)

model = AutoModelForCausalLM.from_pretrained(
    MODEL_PATH,
    torch_dtype=torch.float32,
    device_map="cpu"
)

app = FastAPI()


# =========================================
# REQUEST MODEL
# =========================================

class QuestionRequest(BaseModel):
    sessionType: int
    difficultyLevel: int | None = None
    trackName: int


# =========================================
# HELPER: GENERATE TEXT USING QWEN TEMPLATE
# =========================================

def generate_from_model(prompt: str):

    messages = [
        {"role": "system", "content": "You are a professional interview question generator."},
        {"role": "user", "content": prompt}
    ]

    formatted_prompt = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True
    )

    inputs = tokenizer(formatted_prompt, return_tensors="pt")

    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=1200,
            temperature=0.7
        )

    decoded = tokenizer.decode(outputs[0], skip_special_tokens=True)

    return decoded


# =========================================
# PARSE Q/A FORMAT
# =========================================

def parse_qa_blocks(text: str):

    blocks = text.split("\n\n")
    results = []

    for block in blocks:
        if "Q:" in block and "A:" in block:
            parts = block.split("A:")
            question = parts[0].replace("Q:", "").strip()
            answer = parts[1].strip()
            results.append((question, answer))

    return results




# =========================================
# MAIN ENDPOINT
# =========================================

@app.post("/generate-questions")
def generate_questions(request: QuestionRequest):

    if request.sessionType not in SESSION_TYPES:
        raise HTTPException(status_code=400, detail="Invalid session type")

    session_type = SESSION_TYPES[request.sessionType]

    # ---------------- SOFT SKILLS ----------------

    if session_type == "softskills":

        prompt = """
Generate 10 behavioral interview questions.
Format exactly as:
Q: ...
A: ...
"""

    # ---------------- TECHNICAL ----------------

    elif session_type == "technical":

        if request.trackName not in TRACKS:
            raise HTTPException(status_code=400, detail="Track not supported")

        difficulty = request.difficultyLevel or 1

        prompt = f"""
Generate 10 General Programming interview questions.
Difficulty level: {difficulty}
Format exactly as:
Q: ...
A: ...
"""

    else:
        raise HTTPException(status_code=400, detail="Invalid session type")

    # -------- Generate once --------
    raw_output = generate_from_model(prompt)

    qa_pairs = parse_qa_blocks(raw_output)

    if len(qa_pairs) == 0:
        raise HTTPException(status_code=500, detail="Model failed to generate valid Q/A format")

    response = []

    for idx, (question, answer) in enumerate(qa_pairs[:10], 1):

        response.append({
            "questionText": question,
            "questionId": idx,
            "questionIdealAnswer": answer
        })

    return response