Chatbot RAG com Pinecone e OpenAI

Bancos de dados vetoriais para incorporações com Pinecone

James Chapman

Curriculum Manager, DataCamp

Retrieval Augmented Generation (RAG)

Gerar embedding da consulta
Recuperar documentos similares
Adicionar documentos ao prompt

geração aumentada por recuperação

Inicializar Pinecone e OpenAI

from openai import OpenAI
from pinecone import Pinecone
import pandas as pd
from uuid import uuid4


client = OpenAI(api_key="OPENAI_API_KEY")
pc = Pinecone(api_key="PINECONE_API_KEY")


index = pc.Index("semantic-search-datacamp")

Transcrições do YouTube

youtube_df = pd.read_csv('youtube_rag_data.csv')

| id | blob | channel_id | end | published | start | text | title | url |
|----|------|------------|-----|-----------|-------|------|-------|-----|
|int | dict | str        | int | datetime  | int   | str  | str   | str |

Ingestão de documentos

batch_limit = 100


for batch in np.array_split(youtube_df, len(youtube_df) / batch_limit):

    metadatas = [{"text_id": row['id'], "text": row['text'],
      "title": row['title'], "url": row['url'], 
      "published": row['published']} for  _, row in batch.iterrows()]

    texts = batch['text'].tolist()

    ids = [str(uuid4()) for _ in range(len(texts))]


    response = client.embeddings.create(input=texts, model="text-embedding-3-small")
    embeds = [np.array(x.embedding) for x in response.data]


    index.upsert(vectors=zip(ids, embeds, metadatas), namespace='youtube_rag_dataset')

Função de recuperação

def retrieve(query, top_k, namespace, emb_model):

    query_response = client.embeddings.create(input=query, model=emb_model)
    query_emb = query_response.data[0].embedding


    retrieved_docs = []
    sources = []
    docs = index.query(vector=query_emb, top_k=top_k,
                       namespace='youtube_rag_dataset', include_metadata=True)


    for doc in docs['matches']:
        retrieved_docs.append(doc['metadata']['text'])
        sources.append((doc['metadata']['title'], doc['metadata']['url']))

    return retrieved_docs, sources

Saída da recuperação

query = "How to build next-level Q&A with OpenAI"
documents, sources = retrieve(query, top_k=3, namespace='youtube_rag_dataset',
                              emb_model="text-embedding-3-small")

Document: To use for Open Domain Question Answering. We're going to start...
Source: How to build a Q&A AI in Python [...], https://youtu.be/w1dMEWm7jBc

Document: Over here we have Google and we can ask Google questions...
Source: How to build next-level Q&A with OpenAI, https://youtu.be/coaaSxys5so

Document: We need vector database to enhance the quality of Q&A systems...
Source: How to Build Custom Q&A Transfo [...], https://youtu.be/ZIRmXKHp0-c

Função de criação de prompt com contexto

def prompt_with_context_builder(query, docs):
    delim = '\n\n---\n\n'
    prompt_start = 'Answer the question based on the context below.\n\nContext:\n'
    prompt_end = f'\n\nQuestion: {query}\nAnswer:'

    prompt = prompt_start + delim.join(docs) + prompt_end
    return prompt

Saída do prompt com contexto

query = "How to build next-level Q&A with OpenAI"
context_prompt = prompt_with_context_builder(query, documents)

A saída da função prompt_with_context_builder

Função de perguntas e respostas

def question_answering(prompt, sources, chat_model):

    sys_prompt = "You are a helpful assistant that always answers questions."

    res = client.chat.completions.create(
        model=chat_model,
        messages=[{"role": "system", "content": sys_prompt},
            {"role": "user", "content": prompt}
        ], temperature=0)

    answer = res.choices[0].message.content.strip()
    answer += "\n\nSources:"
    for source in sources:
        answer += "\n" + source[0] + ": " + source[1]
    return answer

Saída de perguntas e respostas

query = "How to build next-level Q&A with OpenAI"
answer = question_answering(prompt_with_context, sources,
                            chat_model='gpt-4o-mini')

A saída da função question_answering

Juntando tudo

query = "How to build next-level Q&A with OpenAI"
documents, sources = retrieve(query, top_k=3, 
                              namespace='youtube_rag_dataset', 
                              emb_model="text-embedding-3-small")
prompt_with_context = prompt_with_context_builder(query, documents)
answer = question_answering(prompt_with_context, sources,
                            chat_model='gpt-4o-mini')

Vamos praticar!

Bancos de dados vetoriais para incorporações com Pinecone