app.py

import os
import json
from typing import Any, Dict, List, Optional, Union

from fastapi import FastAPI, HTTPException
from pydantic import BaseModel, Field
from huggingface_hub import hf_hub_download
from llama_cpp import Llama

app = FastAPI(title="OpenFable API — SynastrIA Networks")

REPO_ID = "SynastriaNetworks/OpenFable-4B-GGUF"

# ── Mapa de modelos disponíveis: apelido -> arquivo GGUF no repo ──────────────
AVAILABLE_MODELS: Dict[str, str] = {
    "openfable-4b":      "openfable-4b-q4_k_m.gguf",  # padrão (melhor custo-benefício)
    # "openfable-4b-q8": "openfable-4b-q8_0.gguf",    # adicione quando publicar
}
DEFAULT_MODEL = "openfable-4b"

# Cache lazy: 1 modelo por vez em RAM (HF Spaces free tem ~16 GB mas é CPU-only)
_loaded: Dict[str, Llama] = {}


def get_model(name: str) -> Llama:
    if name not in AVAILABLE_MODELS:
        raise HTTPException(
            status_code=404,
            detail=f"Modelo '{name}' não encontrado. Disponíveis: {list(AVAILABLE_MODELS)}"
        )
    if name not in _loaded:
        # Evicta o modelo anterior pra liberar RAM
        for k in list(_loaded.keys()):
            print(f"[MODEL] Evictando {k}...")
            del _loaded[k]

        filename = AVAILABLE_MODELS[name]
        print(f"[MODEL] Baixando {filename}...")
        path = hf_hub_download(repo_id=REPO_ID, filename=filename)
        print(f"[MODEL] Carregando {name}...")
        _loaded[name] = Llama(
            model_path=path,
            n_ctx=2048,
            n_threads=2,   # HF Spaces CPU — mais de 1 thread causa GGML_ASSERT
            n_batch=512,
            verbose=False,
            chat_format="chatml",
        )
        print(f"[MODEL] {name} pronto!")
    return _loaded[name]


# Pré-carrega o modelo padrão no boot
print(f"[STARTUP] Carregando modelo padrão: {DEFAULT_MODEL}")
get_model(DEFAULT_MODEL)
print("[STARTUP] Pronto!")


# ── Schemas ───────────────────────────────────────────────────────────────────

class GenerateRequest(BaseModel):
    """Endpoint legado — retrocompatibilidade com o Lucian backend."""
    prompt: str
    model: str = DEFAULT_MODEL
    max_tokens: int = 512
    temperature: float = 0.5


class ToolFunction(BaseModel):
    name: str
    description: str
    parameters: Dict[str, Any] = Field(default_factory=dict)


class Tool(BaseModel):
    type: str = "function"
    function: ToolFunction


class ChatMessage(BaseModel):
    role: str                                          # system | user | assistant | tool
    content: Optional[str] = None
    tool_calls: Optional[List[Dict[str, Any]]] = None  # resposta do modelo com chamadas
    tool_call_id: Optional[str] = None                 # resultado de uma tool call
    name: Optional[str] = None                         # nome da tool (role=tool)


class ChatCompletionRequest(BaseModel):
    model: str = DEFAULT_MODEL
    messages: List[ChatMessage]
    tools: Optional[List[Tool]] = None
    tool_choice: Union[str, Dict[str, Any]] = "auto"
    max_tokens: int = 512
    temperature: float = 0.5


# ── Endpoints ─────────────────────────────────────────────────────────────────

@app.get("/")
def health():
    return {
        "status": "healthy",
        "default_model": DEFAULT_MODEL,
        "available_models": list(AVAILABLE_MODELS.keys()),
    }


@app.get("/v1/models")
def list_models():
    """Lista modelos disponíveis no formato OpenAI."""
    return {
        "object": "list",
        "data": [
            {
                "id": name,
                "object": "model",
                "owned_by": "SynastriaNetworks",
                "file": AVAILABLE_MODELS[name],
            }
            for name in AVAILABLE_MODELS
        ],
    }


@app.post("/v1/generate")
def generate(req: GenerateRequest):
    """Legado: completion por prompt cru, sem histórico de chat."""
    llm = get_model(req.model)
    try:
        out = llm(
            prompt=req.prompt,
            max_tokens=req.max_tokens,
            temperature=req.temperature,
        )
        return {"response": out["choices"][0]["text"], "model": req.model}
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))


@app.post("/v1/chat/completions")
def chat_completions(req: ChatCompletionRequest):
    """
    Endpoint principal — compatível com a API OpenAI.
    Suporta:
      - Histórico de mensagens (system / user / assistant)
      - Tool calling (tools + tool_choice)
      - Seleção de modelo via campo 'model'
    """
    llm = get_model(req.model)

    # Converte os schemas Pydantic para dicts simples que o llama-cpp-python espera
    messages: List[Dict[str, Any]] = []
    for m in req.messages:
        entry: Dict[str, Any] = {"role": m.role}
        if m.content is not None:
            entry["content"] = m.content
        if m.tool_calls:
            entry["tool_calls"] = m.tool_calls
        if m.tool_call_id:
            entry["tool_call_id"] = m.tool_call_id
        if m.name:
            entry["name"] = m.name
        messages.append(entry)

    kwargs: Dict[str, Any] = {
        "messages": messages,
        "max_tokens": req.max_tokens,
        "temperature": req.temperature,
    }

    if req.tools:
        kwargs["tools"] = [
            {
                "type": t.type,
                "function": {
                    "name": t.function.name,
                    "description": t.function.description,
                    "parameters": t.function.parameters,
                },
            }
            for t in req.tools
        ]
        kwargs["tool_choice"] = req.tool_choice

    try:
        # create_chat_completion retorna um dict no formato OpenAI — repassa direto
        response = llm.create_chat_completion(**kwargs)
        return response
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))