What is LlamaIndex? Exploring LLM Orchestration Frameworks

What is LlamaIndex?

LlamaIndex is an orchestration framework for large language model application that simplifies integrating private and public data. It provides tools for data ingestion, indexing, and querying, making it a versatile solution for generative AI needs.

With generative AI rapidly integrating into application development processes, there is an increasing need to integrate private data with public training data of large language models (LLMs). However, most private data is unstructured, siloed, and not in a format that can be readily accessible by LLMs.

In a recent webinar on large language models for the Enterprise, we explored uses for LLMs beyond ChatGPT; we saw how LLM apps need to augment their publicly available training data with private data. This is whereLlamaIndex comes into play, providing an orchestration framework for building LLM apps using built-in tools to ingest and query private data.

Using LlamaIndex as a framework for data integration

At the heart of all generative AI functionality is data. Enterprise applications need to be able to access more than just the public data that LLMs are trained on and need to incorporate structured, unstructured, and semi-structured data from all their internal and external data sources for building applications.

LlamaIndex provides this data integration by using data from multiple unique sources, embedding that data as vectors, and storing that vectorized data in a vector database. Ii then uses that data to perform complex operations like vector search with low latency response times.

Benefits of LlamaIndex

• Simplified data ingestion connecting existing data sources like APIs, PDFs, SQL, NoSQL, documents, etc.
• Store and index private data for different application use cases, with native integration with downstream vector store/vector databases.
• Built-in query interface, providing the ability to return knowledge-augmented responses from input prompts on your data.

How does LlamaIndex work?

LlamaIndex, formerly known as GPT Index, is a framework that provides the tools needed to manage the end-to-end lifecycle for building LLM-based applications. The challenge with building LLM-based applications is that they need data, typically from multiple different sources, and unless there is strong adherence to a common data representation the data required is in many different formats, some highly structured, some unstructured, and some in between.

That is where LlamaIndex provides the toolbox to unlock this data with tools for data ingestion and data indexing. Once ingested and indexed, retrieval augmented generation (RAG) applications can use the LlamaIndex query interface for accessing that data and powering LLMs.

Ingestion

LlamaIndex has 100s of data loaders that provide the ability to connect custom data sources to LLMs. It connects pre-built solutions like Airtable, Jira, and Salesforce to generic plugins for loading data from files, JSON documents, simple CSV, and unstructured data.

A complete list of data loaders can be found on the Llama Hub.

Indexing

Once data is ingested, that data needs to be mathematically represented for LLM querying. With LlamaIndex, an index provides the ability to mathematically represent data in multiple dimensions. Indexing data isn’t a new concept, however, with machine learning, we can expand the granularity of indexing from one or two dimensions (key/value representation for example) to hundreds or thousands of dimensions.

The most common approach to indexing data for machine learning and LLMs is called a vector index; once data has been indexed the mathematical representation of the data is called a vector embedding. There are many types of indexing and embedding models but once data has been embedded the mathematical representation of the data can be used for semantic search – vector-embedded texts with similar meanings will have a similar mathematical representation. For example, king and queen might be highly related if the query is royalty but not highly related if the query is gender.

Querying

This is where the real power of LlamaIndex and LLMs comes into play. Querying data using LlamaIndex isn’t a complex series of commands to merge/join and find the data; it is represented as natural language via a concept called prompt engineering. The simplest way to view interaction with your data once you have ingested and indexed it is that querying becomes a process of asking questions and getting responses.

What are the different indexes in LlamaIndex?

LlamaIndex offers several indexing models to optimize the exploration and categorization of your data. If you know the type of operations your application will perform on data, leveraging a specific index type can provide significant performance gains to the application using the LLM and instantiating the query.

List index

A list index breaks down the data and represents the data as a sequential list. This has the advantage of the data being explorable in a multidimensional manner. This index type works well with structured objects that occur over time – things like change logs where you query how things have changed over time.

Tree index

In a tree index, data is organized as parent and leaf nodes. A tree index lets you traverse large amounts of data and construct responses based on how the search traverses the tree. Tree indexing works best for cases with a pattern of information you want to follow or validate, like building a natural language processing chatbot on top of a support/FAQ engine.

Vector store index

When using the vector store index type, LlamaIndex stores data notes as vector embeddings. This is the most common indexing type as it provides the ability to use the representation of the data in multiple different ways including vector or similarity search. When data is indexed with a vector store index, it can be leveraged locally for smaller datasets and by a single application or for larger datasets and/or to be used across multiple different LLMs/applications it can be stored in a high-performance vector database like Astra DB.

Keyword index

Keyword indexing is more of the traditional approach of mapping a metadata tag, i.e., a keyword, to specific nodes that contain those keywords. Since a keyword may map to multiple different nodes, this mapping builds a web of keyword relationships. This indexing model works well if you tag large volumes of data and query it based on specific keywords across multiple datasets – for example, legal briefings, medical records, or any other data that needs to be aligned based on specific types of metadata.