backend.py

from langchain.chains import RetrievalQA
from langchain_community.llms import Ollama
from langchain_community.vectorstores import FAISS
from langchain_community.embeddings import HuggingFaceEmbeddings
from langchain.prompts import PromptTemplate

INDEX_PATH = "data/index"

QA_PROMPT = PromptTemplate(
    input_variables=["context", "question"],
    template="""
Você é um assistente de compliance. Responda com base apenas no contexto fornecido.
Se não souber, diga: "Desculpe, não encontrei essa informação nos documentos."

Contexto:
{context}

Pergunta:
{question}
"""
)

def ask_question(question):
    embedding = HuggingFaceEmbeddings(
        model_name="sentence-transformers/paraphrase-albert-small-v2",
        encode_kwargs={"normalize_embeddings": True}
    )
    db = FAISS.load_local(INDEX_PATH, embedding, allow_dangerous_deserialization=True)
    retriever = db.as_retriever()

    llm = Ollama(model="tinyllama", temperature=0.3)

    qa = RetrievalQA.from_chain_type(
        llm=llm,
        retriever=retriever,
        return_source_documents=False,
        chain_type_kwargs={"prompt": QA_PROMPT}
    )

    result = qa.invoke({"query": question})
    return result["result"]