“Turn your enterprise data into production-ready LLM applications,” blares the LlamaIndex home page in 60 point type. OK, then. The subhead for that is “LlamaIndex is the leading data framework for building LLM applications.” I’m not so sure that it’s the leading data framework, but I’d certainly agree that it’s a leading data framework for building with large language models, along with LangChain and Semantic Kernel, about which more later.
LlamaIndex currently offers two open source frameworks and a cloud. One framework is in Python; the other is in TypeScript. LlamaCloud (currently in private preview) offers storage, retrieval, links to data sources via LlamaHub, and a paid proprietary parsing service for complex documents, LlamaParse, which is also available as a stand-alone service.
LlamaIndex boasts strengths in loading data, storing and indexing your data, querying by orchestrating LLM workflows, and evaluating the performance of your LLM application. LlamaIndex integrates with over 40 vector stores, over 40 LLMs, and over 160 data sources. The LlamaIndex Python repository has over 30K stars.
Typical LlamaIndex applications perform Q&A, structured extraction, chat, or semantic search, and/or serve as agents. They may use retrieval-augmented generation (RAG) to ground LLMs with specific sources, often sources that weren’t included in the models’ original training.
LlamaIndex competes with LangChain, Semantic Kernel, and Haystack. Not all of these have exactly the same scope and capabilities, but as far as popularity goes, LangChain’s Python repository has over 80K stars, almost three times that of LlamaIndex (over 30K stars), while the much newer Semantic Kernel has over 18K stars, a little over half that of LlamaIndex, and Haystack’s repo has over 13K stars.
Repository age is relevant because stars accumulate over time; that’s also why I qualify the numbers with “over.” Stars on GitHub repos are loosely correlated with historical popularity.
LlamaIndex, LangChain, and Haystack all boast a number of major companies as users, some of whom use more than one of these frameworks. Semantic Kernel is from Microsoft, which doesn’t usually bother publicizing its users except for case studies.
LlamaIndex features
At a high level, LlamaIndex is designed to help you build context-augmented LLM applications, which basically means that you combine your own data with a large language model. Examples of context-augmented LLM applications include question-answering chatbots, document understanding and extraction, and autonomous agents.
The tools that LlamaIndex provides perform data loading, data indexing and storage, querying your data with LLMs, and evaluating the performance of your LLM applications:
- Data connectors ingest your existing data from their native source and format.
- Data indexes, also called embeddings, structure your data in intermediate representations.
- Engines provide natural language access to your data. These include query engines for question answering, and chat engines for multi-message conversations about your data.
- Agents are LLM-powered knowledge workers augmented by software tools.
- Observability/Evaluation integrations enable you to experiment, evaluate, and monitor your app.
Context augmentation
LLMs have been trained on large bodies of text, but not necessarily text about your domain. There are three major ways to perform context augmentation and add information about your domain, supplying documents, doing RAG, and fine-tuning the model.
The simplest context augmentation method is to supply documents to the model along with your query, and for that you might not need LlamaIndex. Supplying documents works fine unless the total size of the documents is larger than the context window of the model you’re using, which was a common issue until recently. Now there are LLMs with million-token context windows, which allow you to avoid going on to the next steps for many tasks. If you plan to perform many queries against a million-token corpus, you’ll want to cache the documents, but that’s a subject for another time.
Retrieval-augmented generation combines context with LLMs at inference time, typically with a vector database. RAG procedures often use embedding to limit the length and improve the relevance of the retrieved context, which both gets around context window limits and increases the probability that the model will see the information it needs to answer your question.
Essentially, an embedding function takes a word or phrase and maps it to a vector of floating point numbers; these are typically stored in a database that supports a vector search index. The retrieval step then uses a semantic similarity search, often using the cosine of the angle between the query’s embedding and the stored vectors, to find “nearby” information to use in the augmented prompt.
Fine-tuning LLMs is a supervised learning process that involves adjusting the model’s parameters to a specific task. It’s done by training the model on a smaller, task-specific or domain-specific data set that’s labeled with examples relevant to the target task. Fine-tuning often takes hours or days using many server-level GPUs and requires hundreds or thousands of tagged exemplars.
Installing LlamaIndex
You can install the Python version of LlamaIndex three ways: from the source code in the GitHub repository, using the llama-index
starter install, or using llama-index-core
plus selected integrations. The starter installation would look like this:
pip install llama-index
This pulls in OpenAI LLMs and embeddings in addition to the LlamaIndex core. You’ll need to supply your OpenAI API key (see here) before you can run examples that use it. The LlamaIndex starter example is quite straightforward, essentially five lines of code after a couple of simple setup steps. There are many more examples in the repo, with documentation.
Doing the custom installation might look something like this:
pip install llama-index-core llama-index-readers-file llama-index-llms-ollama llama-index-embeddings-huggingface
That installs an interface to Ollama and Hugging Face embeddings. There’s a local starter example that goes with this installation. No matter which way you start, you can always add more interface modules with pip
.
If you prefer to write your code in JavaScript or TypeScript, use LlamaIndex.TS (repo). One advantage of the TypeScript version is that you can run the examples online on StackBlitz without any local setup. You’ll still need to supply an OpenAI API key.
LlamaCloud and LlamaParse
LlamaCloud is a cloud service that allows you to upload, parse, and index documents and search them using LlamaIndex. It’s in a private alpha stage, and I was unable to get access to it. LlamaParse is a component of LlamaCloud that allows you to parse PDFs into structured data. It’s available via a REST API, a Python package, and a web UI. It is currently in a public beta. You can sign up to use LlamaParse for a small usage-based fee after the first 7K pages a week. The example given comparing LlamaParse and PyPDF for the Apple 10K filing is impressive, but I didn’t test this myself.
LlamaHub
LlamaHub gives you access to a large collection of integrations for LlamaIndex. These include agents, callbacks, data loaders, embeddings, and about 17 other categories. In general, the integrations are in the LlamaIndex repository, PyPI, and NPM, and can be loaded with pip install
or npm install
.
create-llama CLI
create-llama is a command-line tool that generates LlamaIndex applications. It’s a fast way to get started with LlamaIndex. The generated application has a Next.js powered front end and a choice of three back ends.
RAG CLI
RAG CLI is a command-line tool for chatting with an LLM about files you have saved locally on your computer. This is only one of many use cases for LlamaIndex, but it’s quite common.
LlamaIndex components
The LlamaIndex Component Guides give you specific help for the various parts of LlamaIndex. The first screenshot below shows the component guide menu. The second shows the component guide for prompts, scrolled to a section about customizing prompts.
Learning LlamaIndex
Once you’ve read, understood, and run the starter example in your preferred programming language (Python or TypeScript), I suggest that you read, understand, and try as many of the other examples as look interesting. The screenshot below shows the result of generating a file called essay by running essay.ts and then asking questions about it using chatEngine.ts. This is an example of using RAG for Q&A.
The chatEngine.ts program uses the ContextChatEngine, Document, Settings, and VectorStoreIndex components of LlamaIndex. When I looked at the source code, I saw that it relied on the OpenAI gpt-3.5-turbo-16k model; that may change over time. The VectorStoreIndex module seemed to be using the open-source, Rust-based Qdrant vector database, if I was reading the documentation correctly.
Bringing context to LLMs
As you’ve seen, LlamaIndex is fairly easy to use to create LLM applications. I was able to test it against OpenAI LLMs and a file data source for a RAG Q&A application with no issues. As a reminder, LlamaIndex integrates with over 40 vector stores, over 40 LLMs, and over 160 data sources; it works for several use cases, including Q&A, structured extraction, chat, semantic search, and agents.
I’d suggest evaluating LlamaIndex along with LangChain, Semantic Kernel, and Haystack. It’s likely that one or more of them will meet your needs. I can’t recommend one over the others in a general way, as different applications have different requirements.
Pros
- Helps to create LLM applications for Q&A, structured extraction, chat, semantic search, and agents
- Supports Python and TypeScript
- Frameworks are free and open source
- Lots of examples and integrations
Cons
- Cloud is limited to private preview
- Marketing is slightly overblown
Cost
Open source: free. LlamaParse import service: 7K pages per week free, then $3 per 1000 pages.
Platform
Python and TypeScript, plus cloud SaaS (currently in private preview).
Copyright © 2024 IDG Communications, Inc.