app.py

import os
import faiss
import pickle
import gradio as gr
import spaces
import uvicorn
import threading

from fastapi import FastAPI
from pydantic import BaseModel
from sentence_transformers import SentenceTransformer
from huggingface_hub import hf_hub_download
from openai import OpenAI

# ===============================
# CONFIG
# ===============================

HF_TOKEN = os.environ.get("HF_TOKEN")

# OpenAI-compatible Hugging Face client
client = OpenAI(
    base_url="https://router.huggingface.co/v1",
    api_key=HF_TOKEN
)

MODEL_ID = "meta-llama/Llama-3.1-8B-Instruct:fastest"

# ===============================
# LOAD EMBEDDINGS
# ===============================

print("Loading embedding model...")
embedding_model = SentenceTransformer("all-MiniLM-L6-v2")

# ===============================
# LOAD RAG DATA
# ===============================

print("Downloading FAISS index...")
index_path = hf_hub_download(
    repo_id="mahmodGendy/startup-llama-model",
    filename="faiss.index"
)
index = faiss.read_index(index_path)

print("Downloading documents...")
docs_path = hf_hub_download(
    repo_id="mahmodGendy/startup-llama-model",
    filename="docs.pkl"
)
documents = pickle.load(open(docs_path, "rb"))

print("RAG system ready.")

# ===============================
# RAG RETRIEVAL
# ===============================

def retrieve_context(query, top_k=5):
    query_embedding = embedding_model.encode([query])
    D, I = index.search(query_embedding, top_k)
    retrieved_docs = [documents[i] for i in I[0]]
    return "\n".join(retrieved_docs)

# ===============================
# GPU / Hosted Inference
# ===============================

@spaces.GPU
def ask_llama(user_input):
    context = retrieve_context(user_input)

    evaluation_keywords = [
        "idea", "start", "business",
        "startup", "viable", "launch"
    ]

    is_eval = any(w in user_input.lower() for w in evaluation_keywords)

    if is_eval:
        response_style = """
1. Problem Validation
2. Market Evaluation
3. Risks
4. Improvement Suggestions
"""
    else:
        response_style = "Respond naturally and conversationally."

    system_prompt = f"""
You are a startup validation expert.

Language Rule:
- English → English
- MSA Arabic → MSA Arabic
- Egyptian dialect → Egyptian Arabic

Context:
{context}

{response_style}
"""

    messages = [
        {"role": "system", "content": system_prompt},
        {"role": "user", "content": user_input}
    ]

    # Use hosted API: Hugging Face OpenAI-compatible
    completion = client.chat.completions.create(
        model=MODEL_ID,
        messages=messages,
        max_tokens=400,
        temperature=0.7,
        top_p=0.9
    )

    return completion.choices[0].message.content

# ===============================
# FASTAPI
# ===============================

app = FastAPI()

class Query(BaseModel):
    question: str

@app.post("/ask")
def ask(query: Query):
    answer = ask_llama(query.question)
    return {"answer": answer}

# ===============================
# GRADIO (Required for ZeroGPU)
# ===============================

def gradio_wrapper(question):
    return ask_llama(question)

demo = gr.Interface(
    fn=gradio_wrapper,
    inputs=gr.Textbox(label="Ask your startup question"),
    outputs=gr.Textbox(label="Response")
)

# ===============================
# START SERVERS
# ===============================

def run_fastapi():
    uvicorn.run(app, host="0.0.0.0", port=8000)

if __name__ == "__main__":
    # Start FastAPI in background
    threading.Thread(target=run_fastapi).start()

    # Start Gradio (required for ZeroGPU detection)
    demo.launch(server_name="0.0.0.0", server_port=7860)