LlamaIndex Newsletter 2024-03-05

Greetings, LlamaIndex devotees! 🦙

It was another fun week to be at the center of the LLM universe, and we have tons to share!

🤩 The highlights:

• We shared our thoughts on the future of long-context RAG. As LLMs with context windows over 1M tokens begin to appear, what changes about RAG, and how will LlamaIndex evolve? Tweet, Blog post
• llama-index-networks lets you build a super-RAG application by combining answers from independent RAG apps over the network. Tweet, Blog post, repo
• People loved our release of LlamaParse, a world-beating PDF parsing service, so we made it even better! Tweet, blog post

✨ Feature Releases and Enhancements:

• We released a new llama-index-networks feature that lets you combine multiple independent RAG applications over the network, allowing you to run a single query across all the applications and get a single, combined answer. Tweet, Blog post, repo
• Inference engine Groq wowed us and the world with their incredibly fast query times and we were delighted to introduce first-class support for their LLM APIs. Tweet, notebook
• Users love LlamaParse, the world-beating PDF parsing service we released last week. We pushed improved parsing and OCR support for 81+ languages! We also increased the usage cap from 1k to 10k pages per day. Tweet, blog post
• We migrated our blog off of Medium, we hope you like the new look and the absence of nag screens!
• RAPTOR is a new tree-structured technique for advanced RAG; we turned the paper into a LlamaPack, allowing you to use the new technique in one line of code. Tweet, package, notebook, original paper

🎥 Demos:

• The Koda Retriever is a new retrieval concept: hybrid search where the alpha parameter controlling the importance of vector search vs. keyword search is tuned on a per-query basis by the LLM itself, based on a few-shot examples. Tweet, notebook, package, blog post
• Mixedbread.ai released some state-of-the-art rerankers that perform better than anything seen before; we whipped up a quick cookbook to show you how to use them directly in LlamaIndex. Tweet, Notebook, blog post

🗺️ Guides:

• Function-calling cookbook with open source models shows you how to use Fireworks AI’s OpenAI-compatible API to use all native LlamaIndex support for function calling. Notebook, Tweet.
• We released a best practices cookbook showing how to use LlamaParse, our amazing PDF parser. Tweet, notebook
• A comprehensive guide to semantic chunking for RAG by Florian June covers embedding-based chunking, BERT-based chunking techniques, and LLM-based chunking for everything you need to know about this highly effective technique to improve retrieval quality. Tweet, Blog post

✍️ Tutorials:

• Our own Andrei presented a notebook on building Basic RAG with LlamaIndex at Vector Institute’s RAG bootcamp. Tweet, Notebook
• ClickHouse presented an in-depth tutorial using LlamaIndex to query both structured and unstructured data, and built a bot that queries Hacker News to find what people are saying about the most popular technologies. Tweet, blog post
• POLM (Python, OpenAI, LlamaIndex, MongoDB) is a new reference architecture for building RAG applications and MongoDB has a beautiful, step-by-step tutorial for building it out. Tweet, blog post

🎥 Webinar:

• Our CEO Jerry Liu will do a joint webinar with Adam Kamor of Tonic.ai about building fully-local RAG applications with Ollama and Tonic. People love local models! Tweet, Registration page
• Jerry also did a webinar with Traceloop on leveling up your LLM application with observability. Tweet, YouTube
• Our hackathon at the beginning of February was a huge success! Check out this webinar in which we invited the winners to come and talk about their projects. Tweet, YouTube.