Mastering RAG with RAPTOR: A comprehensive guide using LlamaIndex

In today’s dynamic AI landscape, mastering advanced techniques like Retrieval-Augmented Generation (RAG) is crucial for Data Engineers, Data Scientists, and ML Engineers. 

RAG combines information retrieval with natural language generation to enhance AI responses with accuracy and context. However, traditional RAG methods have limitations, which are now being addressed through the innovative technique: RAPTOR (Recursive Abstractive Processing for Tree Organized Retrieval). 

Today, we’ll explore into the mechanics of RAG, its challenges, and its benefits. We’ll then discuss how RAPTOR overcomes traditional RAG challenges and utilizes LlamaIndex to advance practical implementation.

What is RAG?#

RAG is a hybrid approach that combines the strengths of information retrieval and generative models to enhance the quality and relevance of generated text. Unlike traditional models that only use their training data, RAG utilizes additional context to give better responses.

How RAG works#

RAG tackles the limitations of large language models (LLMs) by incorporating external knowledge into their generation process.

Here’s a breakdown of how it works:

• Retrieval: The first step involves gathering relevant information. RAG acts like a skilled researcher when a user presents a question or prompt. It consults a vast knowledge base, which could be the entire internet, a company’s internal documents, or any other source of textual data. This retrieval process ensures that the LLM can access the most up-to-date and potentially relevant information to address the user’s query.

• Augmentation: Imagine feeding the retrieved information directly to the LLM. It might be overwhelming! RAG employs various augmentation techniques to make this knowledge more digestible. These techniques can involve summarizing the key points of the retrieved passages or encoding them in a way the LLM can understand efficiently. This augmentation step improves the raw information so the LLM can use it better.

• Generation: The LLM generates the response with its inherent understanding of language and the augmented knowledge from the retrieval stage. This response can take various forms depending on the user’s intent. It could be a direct answer to a question, a creative text format inspired by a prompt, or any other kind of textual output. By combining its language skills with augmented knowledge, the LLM aims to deliver a response that is not just creative but also grounded in factual accuracy.

Benefits of RAG#

RAG offers several advantages over traditional generative models:

• Enhanced accuracy and relevance: RAG incorporates external information during generation, leading to more accurate and contextually relevant responses. This reduces the risk of factual errors or irrelevant information commonly found in models trained on static data.

• Improved knowledge coverage: Unlike models limited to their training data, RAG can access and leverage up-to-date information from external sources. This expands the model’s knowledge base and ensures responses reflect current information.

• Reduced hallucinations: Generative models can sometimes generate believable but incorrect information (“hallucinations”). RAG, which uses real-world data, helps reduce this problem by promoting factual responses.

• Increased adaptability: RAG models can be tailored to specific domains by incorporating relevant knowledge bases and retrieval techniques. This allows them to excel in areas like customer support (company policies) or legal research (case law).

• Enhanced user trust: RAG’s ability to cite sources builds user trust by demonstrating transparency and accountability. Users can verify the information and dive deeper if desired.

• Cost-effective development: RAG leverages pretrained generative models like GPT-3, GPT-4, Gemini, or Claude, reducing development costs compared to building a model from scratch. Additionally, the focus on retrieval allows for efficient updates by incorporating new information sources.

Challenges with traditional RAG#

RAG provides a strong method, but it faces challenges in some situations:

• Context deficiency with long documents: Dividing long documents into uniform chunks for retrieval (a common practice) disrupts information flow and makes it challenging for the LLM to grasp the overarching context. Essential relationships between concepts spread across chunks might be overlooked, leading to inaccurate or incomplete responses.

• Flat retrieval structure: In standard RAG, all retrieved information is treated equally when generating responses. This method doesn’t recognize that important information could be buried deep within the documents. As a result, the LLM may struggle to prioritize and use that information effectively.

• Limited reasoning and fact-checking: While RAG can access external information, its ability to reason over that information or perform robust fact-checking can be limited. This can lead to outputs that combine factual elements with inconsistencies or illogical connections.

• Bias and fairness: The quality and bias inherent in the retrieved documents can be reflected in the RAG output.

• Interpretability and explainability: Understanding how RAG generates its outputs can be difficult because of the complex interaction between retrieving information and generating responses. This lack of interpretability can make it harder to debug and build trust in the RAG system.

What is RAPTOR?#

RAPTOR (Recursive Abstractive Processing for Tree Organized RetrievalSarthi, Parth, Salman Abdullah, Aditi Tuli, Shubh Khanna, Anna Goldie, and Christopher D. Manning. "Raptor: Recursive abstractive processing for tree-organized retrieval." arXiv preprint arXiv:2401.18059 (2024).) is a new LLM information retrieval approach. It involves building a tree structure from documents, allowing the model to consider information at different levels of detail. This is claimed to be more efficient and context-aware than traditional methods.

How does RAPTOR work?#

The following illustration elaborates on how RAPTOR works:

Here’s a breakdown of the RAPTOR algorithm step-by-step:

1. Preprocessing#

The document is segmented into smaller units like sentences or paragraphs. These units are then converted into dense vector embeddings, numerical representations capturing the document’s semantic meaning. This allows for efficient similarity comparisons during retrieval.

2. Recursive processing#

This iterative core process refines the document representation:

• Clustering: A clustering algorithm, typically based on Gaussian Mixture Models (GMMs), groups similar text chunks together. This helps organize related information for better summarization.

• Model-based summarization: Each cluster is sent to an LLM like GPT-3. The LLM generates a concise and informative summary of the text within the cluster.

• Re-embedding: The summaries created by the LLM are then converted back into numerical representations suitable for further processing.

3. Tree construction#

After multiple rounds of clustering and summarization (controlled recursion depth), a hierarchical tree is built:

• Leaf nodes: Original text chunks form the base of the tree.

• Summary nodes: As you move up the tree, each node represents a concise summary of its children, capturing the essence of the sub-document it represents.

• Hierarchical embeddings: Each node in the tree can also be associated with its own vector embedding, capturing the summarized meaning at that level.

This multi-layered representation, with both textual summaries and vector embeddings, allows for efficient retrieval at various levels of detail.

4. Retrieval (Inference)#

Given a query, RAPTOR employs two primary retrieval mechanisms for navigating the tree and retrieving relevant information:

• Tree traversal retrieval: This approach systematically explores the tree structure, starting from the root node and progressing down the branches.

  • Beam searchBeam search is an algorithm used in natural language processing and other areas involving search problems. It is an approximate search method that keeps track of only the top k best candidates at each step, making it efficient in terms of time and memory. By selecting the most promising paths, it aims to find a high-quality solution without exploring every possibility​. is a common technique used for tree traversal. It considers a limited number of the most promising branches at each level, focusing the search on the most relevant parts of the tree based on the query.

• Collapsed tree retrieval: This simplified approach views the tree as a single layer, directly comparing the query embedding to the vector embeddings of all leaf nodes (original text chunks) and summary nodes. This is suitable for factual, keyword-based queries where specific details are needed.

RAPTOR’s ability to choose the appropriate retrieval mechanism based on query complexity and utilize both textual summaries and vector embeddings empowers it to retrieve information at the optimal level of abstraction, satisfying diverse query needs.

What is LlamaIndex?#

LlamaIndex is a powerful toolkit designed to enhance the capabilities of LLMs by enabling efficient data retrieval and manipulation from various sources. It allows LLMs to perform tasks such as question answering, text summarization, and knowledge base construction more accurately and effectively.

Here are some key features of LlamaIndex:

• Data integration: Connects LLMs with diverse data sources like databases, APIs, and files.

• Efficient retrieval: Optimizes data retrieval processes to ensure quick access to relevant information.

• Custom indexing: Supports the creation of custom indexes tailored to specific tasks and datasets.

• Scalability: Handles large volumes of data, making it suitable for extensive applications.

• Flexible querying: Allows for complex queries to enhance LLMs’ understanding and response generation.

• Ease of use: Provides user-friendly interfaces and tools for seamless integration with existing systems.

Implementation of RAPTOR using LlamaIndex#

This section explains the implementation of RAPTOR using LlamaIndex, and steps are as follows:

1. Setup and configuration#

Setting up our environment and configuring the necessary parameters is crucial before diving into the implementation. This step ensures that we have all the required tools and settings for a smooth implementation process.

Let’s start by creating a configuration file to store sensitive information and adjustable parameters:

Code explanation:

• Line 1: Specifies the API key for accessing OpenAI’s services.

• Lines 2–4: Define the models used: "text-embedding-3-small" for embeddings and "gpt-3.5-turbo" for language tasks.

• Line 5: Sets the size of text chunks to 400 tokens for processing.

• Line 6: Indicates an overlap of 50 tokens between chunks to maintain context.

• Line 7: Specifies retrieving the top 2 most similar chunks during searches.

• Line 8: Uses a tree traversal method for navigating and processing data.

• Line 9: Sets the language model’s randomness level, with a low value for more deterministic responses.

Now, let’s implement the setup code:

Code explanation:

• Lines 1–4: Import necessary libraries for the program to function. yaml is imported to read YAML configuration files, os to interact with the operating system (setting environment variables), logging for structured logging messages and Dict, Any from typing for type hints.

• Lines 7–8: Set up basic logging configuration. logging.basicConfig() configures logging to display messages with an INFO level or higher, formatted with a timestamp, logging level, and message content. logger = logging.getLogger(__name__) creates a logger object specific to the current module.

• Lines 10–21: Define the load_config function, which loads configuration from a YAML file. The function attempts to open and parse the YAML file specified by config_path. It uses a try...except block to handle FileNotFoundError if the file is not found or yaml.YAMLError for errors during parsing. Successful loading returns the configuration as a dictionary (Dict[str, Any]).

• Line 24: Calls the load_config function with a specific path to load the configuration into the config variable.

• Line 27: Sets the OpenAI API key by retrieving openai_api_key from the config dictionary. This sets an environment variable needed for interactions with OpenAI services.

• Line 30: Silently installs required Python packages using pip install -q within a Jupyter Notebook environment.

• Line 33: Silently downloads a PDF file from a specified URL and saves it locally as raptor_paper.pdf.

• Line 35: Logs an informational message using the configured logger indicating that the setup process has been completed without errors. This helps track the procedure’s progress and status.

2. Document loading and vector store setup#

We’ll load the document and set up the vector store in this step. This is crucial in preparing our data for efficient retrieval and processing.

Code explanation:

• Lines 1–4: Import libraries essential for document processing and vector storage. nest_asyncio is imported to enable asynchronous operations in Jupyter Notebooks, SimpleDirectoryReader from llama_index.core is used to read documents from a directory, ChromaVectorStore from llama_index.vector_stores.chroma facilitates storing and retrieving document vectors via ChromaDB, and chromadb provides functionality for interacting with ChromaDB.

• Line 7: Applies the nest_asyncio library to enable asynchronous operations in the current Jupyter Notebook environment. This setup allows concurrent task execution without blocking.

• Lines 9–13: Define the load_document function, which loads a document from a specified file path. It uses SimpleDirectoryReader to read the document specified by file_path, logs a success message using the logger (logger.info()), and returns the loaded document (documents).

• Lines 15–21: Define the setup_vector_store function, which sets up a ChromaDB vector store. It creates a PersistentClient object from chromadb to interact with the database located at db_path. It then retrieves or creates a collection named collection_name using client.get_or_create_collection(). A ChromaVectorStore object is instantiated using the obtained collection, logs a success message, and returns the created vector_store.

• Line 24: Invokes the load_document function with the path ./raptor_paper.pdf to load the RAPTOR paper document content into the documents variable.

• Line 27: Calls setup_vector_store with database path ./raptor_paper_db and collection name raptor to initialize a ChromaDB vector store named "raptor". The resulting vector_store object is set up for storing and retrieving document vectors in ChromaDB.

3. RAPTOR pack configuration#

Now that we have loaded our document and vector store, we’ll configure the RAPTOR pack. This step involves setting up the core components of the RAPTOR system.

Code explanation:

• Lines 1–4: Import components necessary for building and configuring a RAPTOR pack. SentenceSplitter from llama_index.core.node_parser facilitates document segmentation into sentences, OpenAI from llama_index.llms.openai provides an interface for using OpenAI’s LLMs, OpenAIEmbedding from llama_index.embeddings.openai generates document embeddings using OpenAI’s models and RaptorPack from llama_index.packs.raptor represents a pre-configured workflow for RAPTOR.

• Lines 6–18: Define the create_raptor_pack function. It takes three arguments:

  • documents: Represents the loaded documents intended for processing within the RAPTOR pack.

  • vector_store: Refers to the ChromaDB vector store object previously created, used for storing and retrieving document vectors.

  • config: A dictionary containing various configuration settings for customizing the RAPTOR pack.

  • Within this function:

    • RaptorPack(...) initializes a RaptorPack object using the following parameters:

• logger.info("RAPTOR pack created successfully."): Logs an informational message confirming the successful creation of the RAPTOR pack.

• return pack: Returns the initialized RaptorPack object, ready for further use in specific tasks as defined by the configured mode.

• Line 21: Calls the create_raptor_pack function with documents, vector_store, and config as arguments, resulting in instantiating a specific RAPTOR pack configured according to the provided documents, vector store, and configuration settings. This raptor_pack instance is now ready for performing tasks such as document retrieval and processing within the specified environment.

4. Retrieval and query engine#

In this final section, we’ll set up the retriever and query engine, which will allow us to perform queries on our processed document.

Code explanation:

• Lines 1–3: Import necessary components for setting up a retrieval system and processing queries. RaptorRetriever from llama_index.packs.raptor likely implements document retrieval functionality within the RAPTOR framework. RetrieverQueryEngine from llama_index.core.query_engine represents a query engine capable of interacting with retrievers and potentially language models for query processing.

• Lines 5–16: Define the create_raptor_retriever function:

  • Takes vector_store, representing the ChromaDB vector store, and config, a dictionary holding configuration settings.

  • Creates a RaptorRetriever object (retriever) with parameters:

• Logs successful creation of the retriever using logger.info.

• Returns the initialized retriever object.

• Lines 18–25: Define the create_query_engine function:

  • Takes retriever, the RaptorRetriever object, and config, the configuration dictionary.

  • Uses RetrieverQueryEngine.from_args() to instantiate a query engine (query_engine) with:

    • retriever: The previously created retriever object.

    • llm=OpenAI(model=config['models']['llm'], temperature=config['temperature']): Configures an OpenAI LLM for query processing.

  • Logs successful creation of the query engine using logger.info.

  • Returns the initialized query_engine object.

• Lines 27–34: Define the run_multiple_queries function:

  • Takes query_engine, the query engine object, and queries, a list of strings representing user queries.

  • Initializes an empty list results to store query-response pairs.

  • Iterates through each query in queries:

    • Uses query_engine.query(query) to process each query, likely retrieving relevant documents and generating a response using the configured LLM.

    • Appends a tuple (query, str(response)) to results, where response is the generated response converted to a string.

    • Logs successful query processing using logger.info.

  • Returns results, a list of tuples containing the original queries and corresponding responses.

• Lines 37–38: Call create_raptor_retriever with vector_store and config to instantiate a specific RaptorRetriever for use in the subsequent steps.

• Lines 41–45: Define example queries related to the RAPTOR research paper stored in queries, setting up different inquiries about RAPTOR’s capabilities or comparisons.

• Line 47: Calls run_multiple_queries with query_engine and queries to process each query using the configured retrieval and query engine setup.

• Lines 50–53: Iterate through results, printing each query and its corresponding response:

  • Prints the original query with print(f"Query: {query}").

  • Prints the generated response from the query engine with print(f"Response: {response}").

  • Separates each query-response pair with a line of dashes print("-" * 50) for clarity.

Here is the output generated by the RAPTOR code above:

Here’s a comparison table highlighting why RAPTOR is considered superior to traditional RAG methods:

From RAGs to riches#

Now that you know a bit about RAPTOR, we hope you feel better equipped to master RAG and its various techniques.

Are you ready to gain more hands-on skills with RAG?

If so, here are a few courses you may find interesting:

You can also start building with RAG through Educative Projects, which guide you through creating tangible outcomes for your portfolio (without the setup):

• Build an LLM-Powered Wikipedia Chat Assistant with RAG

• Building a Retrieval-Augmented Generation System Using FastAPI

• Build an LLM-powered Chatbot with RAG using LlamaIndex

Happy learning!