app.py

import os
import subprocess
import asyncio
import time
import threading
import gc
import psutil
import torch
from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
import uvicorn
from langchain.prompts import PromptTemplate
from langchain_community.chat_models import ChatOllama
from langchain_community.tools import DuckDuckGoSearchRun
from langchain_community.utilities import DuckDuckGoSearchAPIWrapper
from langchain_core.output_parsers import JsonOutputParser, StrOutputParser
from langgraph.graph import END, StateGraph
from typing_extensions import TypedDict

# Asegúrate de tener la librería Ollama instalada
# pip install ollama langchain langchain_community langgraph

# Configuración de Ollama y su servicio
OLLAMA = os.path.expanduser("~/ollama")

def download_ollama_model(model_name='hf.co/MaziyarPanahi/Llama-3.2-3B-Instruct-uncensored-GGUF:IQ1_S'):
    try:
        if not os.path.exists(OLLAMA):
            print("Ollama no encontrado, descargando...")
            subprocess.run("curl -fsSL https://ollama.com/install.sh | sh", shell=True)
            os.chmod(OLLAMA, 0o755)
        
        print(f"Descargando el modelo: {model_name}")
        subprocess.run(["ollama", "pull", model_name], check=True)
    except subprocess.CalledProcessError as e:
        print(f"Error al descargar el modelo: {e}")
        raise

# Función asíncrona para manejar la descarga del modelo
async def async_download_ollama_model():
    await asyncio.to_thread(download_ollama_model)

# Iniciar el servidor Ollama en un hilo
def ollama_service_thread():
    print("Iniciando el servicio de Ollama")
    subprocess.run("~/ollama serve", shell=True)

# Crear un hilo para iniciar Ollama
OLLAMA_SERVICE_THREAD = threading.Thread(target=ollama_service_thread)
OLLAMA_SERVICE_THREAD.start()

# Esperar a que Ollama esté listo
print("Esperando a que Ollama inicie...")
time.sleep(10)

# Descargar el modelo de Hugging Face si no está disponible
asyncio.run(async_download_ollama_model())

# Crear instancia de FastAPI
app = FastAPI()

# Definición del modelo de datos para recibir las consultas en la API
class QueryRequest(BaseModel):
    query: str

# Definir el modelo de lenguaje de Ollama (sin 'temperature')
local_llm = 'hf.co/MaziyarPanahi/Llama-3.2-3B-Instruct-uncensored-GGUF:IQ1_S'
llama3 = ChatOllama(model=local_llm)

# Definir la herramienta de búsqueda web usando DuckDuckGo
wrapper = DuckDuckGoSearchAPIWrapper(max_results=1)
web_search_tool = DuckDuckGoSearchRun(api_wrapper=wrapper)

# Definición de los prompts para generación y enrutamiento
generate_prompt = PromptTemplate(
    template="""
    <|begin_of_text|>
    <|start_header_id|>system<|end_header_id|> 
    You are an AI assistant for Research Question Tasks, that synthesizes web search results. 
    Strictly use the following pieces of web search context to answer the question. If you don't know the answer, just say that you don't know. 
    Keep the answer concise, but provide all of the details you can in the form of a research report. 
    Only make direct references to material if provided in the context.
    <|eot_id|>
    <|start_header_id|>user<|end_header_id|> 
    Question: {question} 
    Web Search Context: {context} 
    Answer: 
    <|eot_id|> 
    <|start_header_id|>assistant<|end_header_id|>""",
    input_variables=["question", "context"],
)

generate_chain = generate_prompt | llama3 | StrOutputParser()

router_prompt = PromptTemplate(
    template="""
    <|begin_of_text|>
    <|start_header_id|>system<|end_header_id|> 
    You are an expert at routing a user question to either the generation stage or web search. 
    Use the web search for questions that require more context for a better answer, or recent events.
    Otherwise, you can skip and go straight to the generation phase to respond.
    You do not need to be stringent with the keywords in the question related to these topics.
    Give a binary choice 'web_search' or 'generate' based on the question. 
    Return the JSON with a single key 'choice' with no premable or explanation. 
    Question to route: {question} 
    <|eot_id|> 
    <|start_header_id|>assistant<|end_header_id|>""",
    input_variables=["question"],
)

question_router = router_prompt | llama3 | JsonOutputParser()

query_prompt = PromptTemplate(
    template="""
    <|begin_of_text|>
    <|start_header_id|>system<|end_header_id|> 
    You are an expert at crafting web search queries for research questions.
    More often than not, a user will ask a basic question that they wish to learn more about, however it might not be in the best format. 
    Reword their query to be the most effective web search string possible.
    Return the JSON with a single key 'query' with no premable or explanation. 
    Question to transform: {question} 
    <|eot_id|> 
    <|start_header_id|>assistant<|end_header_id|>""",
    input_variables=["question"],
)

query_chain = query_prompt | llama3 | JsonOutputParser()

# Definir el estado del grafo
class GraphState(TypedDict):
    question: str
    generation: str
    search_query: str
    context: str

# Nodos de procesamiento
def generate(state):
    print("Step: Generating Final Response")
    question = state["question"]
    context = state["context"]
    generation = generate_chain.invoke({"context": context, "question": question})
    return {"generation": generation}

def transform_query(state):
    print("Step: Optimizing Query for Web Search")
    question = state['question']
    gen_query = query_chain.invoke({"question": question})
    search_query = gen_query.get("query", "")  # Asegurarnos de que estamos obteniendo la clave correcta
    return {"search_query": search_query}

def web_search(state):
    search_query = state['search_query']
    print(f'Step: Searching the Web for: "{search_query}"')
    try:
        search_result = web_search_tool.invoke(search_query)
        if isinstance(search_result, str):  # Si la respuesta es una cadena, la convertimos en un diccionario
            print(f"Respuesta de búsqueda web es cadena: {search_result}")
            return {"context": search_result}
        elif isinstance(search_result, dict):  # Si es un diccionario, lo usamos directamente
            return {"context": search_result}
        else:
            raise ValueError("Respuesta de búsqueda web no es válida")
    except Exception as e:
        print(f"Web search failed: {e}")
        return None  # Si la búsqueda falla, no devuelve contexto

def route_question(state):
    print("Step: Routing Query")
    question = state['question']
    output = question_router.invoke({"question": question})
    if output.get('choice') == "web_search":
        print("Step: Routing Query to Web Search")
        return "websearch"
    elif output.get('choice') == 'generate':
        print("Step: Routing Query to Generation")
        return "generate"

# Crear el grafo de estado
workflow = StateGraph(GraphState)
workflow.add_node("websearch", web_search)
workflow.add_node("transform_query", transform_query)
workflow.add_node("generate", generate)

workflow.set_conditional_entry_point(
    route_question,
    {
        "websearch": "transform_query",
        "generate": "generate",
    },
)
workflow.add_edge("transform_query", "websearch")
workflow.add_edge("websearch", "generate")
workflow.add_edge("generate", END)

# Compilar el agente
local_agent = workflow.compile()

# Función para ejecutar el agente
def run_agent_parallel(query):
    output = local_agent.invoke({"question": query})
    if "generation" not in output:  # Si la búsqueda web falló y no hubo respuesta de generación
        print("Web search failed, using Ollama model directly.")
        return generate({"question": query, "context": ""})["generation"]  # Generar directamente
    return output['generation']

# Lógica del servidor FastAPI
@app.post("/query")
async def query_endpoint(request: QueryRequest):
    query = request.query
    return {"response": run_agent_parallel(query)}

# Lógica de recursos
def release_resources():
    try:
        torch.cuda.empty_cache()
        gc.collect()
    except Exception as e:
        print(f"Failed to release resources: {e}")

def resource_manager():
    MAX_RAM_PERCENT = 1
    MAX_CPU_PERCENT = 1
    MAX_GPU_PERCENT = 1
    MAX_RAM_MB = 1

    while True:
        try:
            virtual_mem = psutil.virtual_memory()
            current_ram_percent = virtual_mem.percent
            current_ram_mb = virtual_mem.used / (1 * 1)  # Convert to MB

            if current_ram_percent > MAX_RAM_PERCENT or current_ram_mb > MAX_RAM_MB:
                release_resources()

            current_cpu_percent = psutil.cpu_percent()
            if current_cpu_percent > MAX_CPU_PERCENT:
                psutil.Process(os.getpid()).nice()

            if torch.cuda.is_available():
                gpu = torch.cuda.current_device()
                gpu_mem = torch.cuda.memory_percent(gpu)

                if gpu_mem > MAX_GPU_PERCENT:
                    release_resources()

        except Exception as e:
            print(f"Error in resource manager: {e}")

resource_manager()

if __name__ == "__main__":
    uvicorn.run(app, host="0.0.0.0", port=8000)