Práticas recomendadas do RAG

Arquitetura geral

1. Compute documentos e vetores de consulta comopenai/text-embedding-3-large
2. Recupere Top-K de uma loja de vetores (FAISS / Milvus / pgvector)
3. Reclassificar comcohere/rerank-english-v3
4. Coloque os pedaços Top-N no prompt e deixecohere/command-r-plusou Claude geram a resposta final

Incorporação

resp = client.embeddings.create(
    model="openai/text-embedding-3-large",
    input=[chunk.text for chunk in chunks],
)
vectors = [d.embedding for d in resp.data]

Modelo de prompt

You are an expert assistant. Answer based ONLY on context.
If the answer is not in the context, say "I don't know".

[Context]
{retrieved_chunks}

[Question]
{user_question}