app.py

import os
import gradio as gr
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.vectorstores import Chroma
from langchain_huggingface import HuggingFaceEmbeddings
from langchain_community.document_loaders import PyPDFLoader
import requests
from rerankers import Reranker
from langchain_huggingface import HuggingFacePipeline
from langchain import hub
from langchain_core.output_parsers import StrOutputParser

# Configurar USER_AGENT
os.environ["USER_AGENT"] = "MyHFSpace/1.0 (HuggingFace Space for Nutrition Q&A)"

# Paso 1: Descargar el documento PDF
URL = "https://gruposdetrabajo.sefh.es/gefp/images/stories/documentos/4-ATENCION-FARMACEUTICA/Nutricion/Manual_basico_N_clinica_y_Dietetica_Valencia_2012.pdf"
response = requests.get(URL)
with open("Manual_de_nutrición_clinica.pdf", "wb") as f:
    f.write(response.content)

# Paso 2: Inicializar el modelo usando Hugging Face Pipeline
# Usamos gpt2 en lugar de distilbert-base-uncased
llm = HuggingFacePipeline.from_model_id(
    model_id="gpt2",  # Cambiado a gpt2, que es compatible con AutoModelForCausalLM
    task="text-generation",
    pipeline_kwargs={"temperature": 0.7, "top_k": 50, "top_p": 0.95, "max_length": 512}
)
chain = llm | StrOutputParser()

# Cargar y procesar el PDF
loader = PyPDFLoader("Manual_de_nutrición_clinica.pdf")
documents = loader.load()
text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=20)
all_splits = text_splitter.split_documents(documents)

# Crear embeddings y la base de datos vectorial
model_name = "sentence-transformers/paraphrase-multilingual-mpnet-base-v2"
embeddings = HuggingFaceEmbeddings(model_name=model_name)
vectordb = Chroma.from_documents(documents=all_splits, embedding=embeddings, persist_directory="chroma_db")

# Inicializar el reranker
ranker = Reranker("answerdotai/answerai-colbert-small-v1", model_type='colbert')

# Paso 3: Definir la función RAG con reranking
def format_docs(docs):
    return "\n\n".join(doc[0].page_content for doc in docs)

def test_rag_reranking(query, ranker):
    docs = vectordb.similarity_search_with_score(query)
    prompt = hub.pull("rlm/rag-prompt")
    rag_chain = prompt | llm | StrOutputParser()
    context = []
    for doc, score in docs:
        if score < 7:
            doc_details = doc.to_json()['kwargs']
            context.append(doc_details['page_content'])
    if len(context) > 0:
        ranking = ranker.rank(query=query, docs=context)
        useful_context = ranking[0].text
        generation = rag_chain.invoke({"context": useful_context, "question": query})
        return generation
    else:
        return "No tengo información para responder a esta pregunta"

# Paso 4: Crear una interfaz con Gradio
def answer_query(query):
    return test_rag_reranking(query, ranker)

interface = gr.Interface(
    fn=answer_query,
    inputs=gr.Textbox(label="Ingresa tu pregunta sobre nutrición:"),
    outputs=gr.Textbox(label="Respuesta:"),
    title="Respuesta a Preguntas sobre Nutrición",
    description="Haz preguntas sobre nutrición basadas en el Manual Básico de Nutrición Clínica y Dietética."
)

# Lanzar la interfaz
interface.launch()