Advanced RAG Techniques Overview

Recent developments in retrieval-augmented generation have led to three distinct paradigms:

1. Naive RAG
2. Advenced RAG
3. Modular RAG

This article explores these approaches and demonstrates how to implement an advanced RAG pipeline using LlamaIndex with Python. We'll cover three key optimization techniques:

1. Pre-retrieval optimization: Sentence window retrieval
2. Retrieval optimization: Hybrid search
3. Post-retrieval optimization: Re-ranking

Advanced RAG Implementation

Advanced RAG addresses limitations of naive approaches through targeted enhancements in data processing, retrieval, and post-processing stages.

Prerequisites

Required packages:

pip install llama-index
pip install weaviate-client llama-index-vector-stores-weaviate

API configuration:

import os
from dotenv import load_dotenv, find_dotenv

load_dotenv(find_dotenv())

Basic RAG Implementation

Configure models and load data:

from llama_index.embeddings.openai import OpenAIEmbedding
from llama_index.llms.openai import OpenAI
from llama_index.core.settings import Settings

Settings.llm = OpenAI(model="gpt-3.5-turbo", temperature=0.1)
Settings.embed_model = OpenAIEmbedding()

from llama_index.core import SimpleDirectoryReader
documents = SimpleDirectoryReader(
    input_files=["./data/sample_document.txt"]
).load_data()

Parse documents and build index:

from llama_index.core.node_parser import SimpleNodeParser

node_parser = SimpleNodeParser.from_defaults(chunk_size=1024)
nodes = node_parser.get_nodes_from_documents(documents)

import weaviate
client = weaviate.Client(embedded_options=weaviate.embedded.EmbeddedOptions())

from llama_index.core import VectorStoreIndex, StorageContext
from llama_index.vector_stores.weaviate import WeaviateVectorStore

vector_store = WeaviateVectorStore(weaviate_client=client, index_name="DocumentIndex")
storage_context = StorageContext.from_defaults(vector_store=vector_store)
index = VectorStoreIndex(nodes, storage_context=storage_context)

Query execution:

query_engine = index.as_query_engine()
response = query_engine.query("Query about document content")

Advanced RAG Enhancements

Sentence window retrieval:

from llama_index.core.node_parser import SentenceWindowNodeParser

node_parser = SentenceWindowNodeParser.from_defaults(
    window_size=3,
    window_metadata_key="context_window",
    original_text_metadata_key="source_text"
)

from llama_index.core.postprocessor import MetadataReplacementPostProcessor
post_processor = MetadataReplacementPostProcessor(target_metadata_key="context_window")

Hybrid search configuration:

query_engine = index.as_query_engine(
    vector_store_query_mode="hybrid",
    alpha=0.5
)

Re-ranking implementation:

from llama_index.core.postprocessor import SentenceTransformerRerank

reranker = SentenceTransformerRerank(
    top_n=2,
    model="BAAI/bge-reranker-base"
)

query_engine = index.as_query_engine(
    similarity_top_k=6,
    node_postprocessors=[reranker]
)