LlamaIndex Newsletter 2024-02-27

LlamaCloud was released last week, kicking off production-level context enhancement for LLM and RAG applications. Its core consists of LlamaParse, which handles complex documents, and the Managed Ingestion and Retrieval API, which connects to 150+ sources and 40+ storage options. LlamaParse is currently in public preview focused on PDFs, while the API is in private preview for select enterprise partners.

This week's highlights are features that expand search, reranking, and model usage.

• Sub-Document Summaries: A new chunking approach that embeds hierarchical metadata into chunks, aiming to reduce the limitations of naive chunking and enable more accurate RAG retrieval.
• MistralAI Cookbook: A guide covering the latest mistral-large model, which supports near-GPT-4 level reasoning, function calling, and JSON output.
• Gemma Cookbook: Provides an example of using GoogleDeepMind's Gemma together with Ollama to build local RAG with 2B/7B models.
• ColBERT Integration: Enables attaching ColBERT reranking in LlamaIndex with a single line, highlighting processing roughly 100x faster than BERT-based models.
• create-llama: Introduces a flow using LlamaPacks to connect advanced RAG into a full-stack web app with two lines of code.

In the demo section, Counselor Copilot, which helps crisis counselors at the Trevor Project, drew attention. It's a social-impact case that assists counseling work supporting LGBTQ+ youth by providing real-time context, suggested responses, and follow-up actions.

Guides, tutorials, and webinars were also rounded up. There's a wealth of practical material, including an advanced RAG development guide covering complex PDF QA, building a React-based chatbot, a RAG development journey from notebook to microservices, a rundown of RAG pain points, and a production RAG webinar with JasperAI.