app.py

from fastapi import FastAPI
from fastapi.responses import StreamingResponse
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel
from llama_cpp import Llama
from huggingface_hub import hf_hub_download
from supabase import create_client
import os, uvicorn, threading
from contextlib import asynccontextmanager

# =========================
# CONFIG
# =========================
HF_TOKEN = os.getenv("HF_TOKEN")
SUPABASE_URL = os.getenv("SUPABASE_URL")
SUPABASE_KEY = os.getenv("SUPABASE_KEY")

supabase = create_client(SUPABASE_URL, SUPABASE_KEY)

model = None

# =========================
# REQUEST
# =========================
class ChatRequest(BaseModel):
    message: str
    request_id: str
    temperature: float = 0.7

# =========================
# CLEAN OUTPUT
# =========================
def clean_output(text):
    stop_words = [
        "<|eot_id|>",
        "<|end_of_text|>",
        "<|eof|>",
        "Human:",
        "Assistant:",
        "User:"
    ]
    for w in stop_words:
        if w in text:
            text = text.split(w)[0]
    return text.strip()

# =========================
# PROMPT
# =========================
def build_prompt(user_msg):
    return f"""<|begin_of_text|>
<|start_header_id|>system<|end_header_id|>
Your name is Llama and you are a cheerful friendly AI buddy made for voice conversation.
Rules:
- Always refer to yourself as Llama
- Speak naturally like a real voice conversation with a friend
- Use casual spoken language like hey sure yep got it
- Answer in 1 to 2 sentences only
- Keep answer under 30 words
- Do not use symbols
- Do not use abbreviations
- Use digits instead of words
- No new lines
- Output plain text only
<|eot_id|>
<|start_header_id|>user<|end_header_id|>
{user_msg}
<|eot_id|>
<|start_header_id|>assistant<|end_header_id|>
"""

# =========================
# MODEL LOAD
# =========================
def load_model():
    return Llama(
        model_path=hf_hub_download(
            repo_id="Valtry/llama3.2-3b-q4-gguf",
            filename="llama3.2-3b-q4.gguf",
            token=HF_TOKEN,
            cache_dir="/data"
        ),
        n_ctx=2048,
        n_threads=4,
        n_batch=512,
        use_mmap=True,
        use_mlock=True,
        f16_kv=True,
        verbose=False
    )

@asynccontextmanager
async def lifespan(app: FastAPI):
    global model
    model = load_model()
    yield

# =========================
# APP
# =========================
app = FastAPI(lifespan=lifespan)

app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_methods=["*"],
    allow_headers=["*"],
)

# =========================
# SAVE
# =========================
def save_message(role, content, request_id):
    supabase.table("messages").insert({
        "role": role,
        "content": content,
        "request_id": request_id
    }).execute()

# =========================
# SUPABASE UPDATE HELPER
# =========================
def update_message(msg_id, content, status=None):
    data = {"content": content}
    if status:
        data["status"] = status
    try:
        supabase.table("messages").update(data).eq("id", msg_id).execute()
    except Exception as e:
        print(f"Supabase update failed: {e}")


# =========================
# CHAT
# =========================

@app.post("/v1/chat")
async def chat(req: ChatRequest):

    def generate():

        prompt = build_prompt(req.message)

        full_text = ""

        stream = model(
            prompt,
            max_tokens=2048,
            temperature=req.temperature,
            top_p=0.9,
            repeat_penalty=1.15,
            stop=["<|eot_id|>", "<|end_of_text|>", "<|eof|>"],
            stream=True
        )

        # 🔥 STREAM DIRECTLY TO ESP
        for chunk in stream:
            token = chunk["choices"][0]["text"]

            full_text += token

            yield token.replace("\n", " ").replace("\r", "")   # ⚡ direct streaming

        # 🔥 SAVE AFTER COMPLETION
        final = clean_output(full_text)

        save_message("user", req.message, req.request_id)
        save_message("assistant", final, req.request_id)

    return StreamingResponse(generate(), media_type="text/plain")

# =========================
# GET RESPONSE
# =========================
@app.get("/v1/get_response/{request_id}")
def get_response(request_id: str):
    try:
        res = supabase.table("messages") \
            .select("content, status") \
            .eq("role", "assistant") \
            .eq("request_id", request_id) \
            .order("created_at", desc=True) \
            .limit(1) \
            .execute()

        data = res.data

        if data:
            return {
                "response": data[0]["content"],
                "status": data[0]["status"]
            }
        else:
            return {"response": None, "status": "waiting"}

    except Exception as e:
        return {"error": str(e)}

# =========================
# ROOT
# =========================
@app.get("/")
def root():
    return {"status": "LLaMA API running"}

# =========================
# RUN
# =========================
if __name__ == "__main__":
    uvicorn.run("app:app", host="0.0.0.0", port=7860)