rerank rag

rag.py

from pydantic import SecretStr
import requests
from pymupdf import pymupdf
from langchain_text_splitters import RecursiveCharacterTextSplitter

print("# 1. Load e Split do PDF")
file = pymupdf.open('./doc.pdf')
content = ""
content += file[0].get_text("text")

splitter = RecursiveCharacterTextSplitter(chunk_size=400, chunk_overlap=40)
chunks = splitter.split_text(content)

print("# 2. Configuração do Llama Server de Rerank")
url = "http://localhost:8080/v1/reranking"
query = "Nome da pessoa X exemplo."

payload = {
    "model": "qwen3-reranker-0.6b-q8_0.gguf",
    "query": query,
    "documents": chunks,
}

response = requests.post(url, json=payload)

if response.status_code != 200:
  raise Exception(f"Erro {response.status_code}: {response.text}")

print("# 3. Processa e ordena os resultados com base na resposta do servidor")
data = response.json()
# O llama-server retorna uma lista contendo o índice original e o score de relevância
results = data.get("results", [])

# Ordena do maior score para o menor
results_sorted = sorted(
    results, key=lambda x: x["relevance_score"], reverse=True
)

print(f"\n--- Top Resultados (Rerank via Llama Server) para: '{query}' ---\n")
for item in results_sorted[:3]:  # Exibe os 3 melhores
  idx = item["index"]
  score = item["relevance_score"]
  print(f"Index: {idx} | Score Reranker: {score:.4f}")
  print(chunks[idx])
  print("-" * 40)

300/30 splitter better