LlamaIndex Newsletter 2024-05-21

Hello LlamaIndex Community! 🦙

Welcome to another exciting weekly update from LlamaIndex! Last week was a standout in the AI world with significant updates on GPT-4o and Gemini models. We're thrilled to bring you a host of exceptional integration updates, guides, tutorials, webinars, and events.

🤩 The highlights:

• LlamaIndex on Vertex AI (Google Cloud): Introducing the new RAG API powered by advanced LlamaIndex modules on Vertex AI (Google Cloud). Docs, Tweet.
• Enhanced Document Parsing with GPT-4o: Integrated GPT-4o in LlamaParse for superior document parsing. Notebook, Tweet.
• Cookbook on Structured Image Extraction with GPT-4o: Check out our new cookbook on using GPT-4o for Structured Image Extraction. Notebook, Tweet.

✨ Feature Releases and Enhancements:

1. We have launched LlamaIndex on Vertex AI (Google Cloud) to introduce the new RAG API, enhanced by LlamaIndex's advanced modules. This integration simplifies setup and enhances user access for the developers with the flexibility to connect a variety of data sources and file types. It fully supports the latest LLMs, including Gemini 1.5 Flash, Gemini 1.5 Pro, and Gemini 1.0 models. Docs, Tweet.
2. We have introduced GPT-4o with LlamaParse, offering enhanced document parsing into markdown for complex files, ensuring higher data quality for your RAG pipeline. Note the increased cost of $0.60 USD per page. Note the increased cost of $0.60 USD per page compared to the standard $0.003 per page. Notebook, Tweet.
3. We have released a cookbook on using GPT-4o for Structured Image Extraction, showing how to convert images into structured JSONs with a 0% failure rate and higher quality than GPT-4V. Notebook, Tweet.
4. LlamaParse integration with Quivr to enhance document parsing capabilities. Now, you can easily process complex documents like PDFs, PPTX, and Markdown files, ensuring clean data storage and accurate retrieval in your personalized AI assistants. Docs, Tweet.

🗺️ Guides:

• Guide to Enhanced QA with LlamaParse on complex tables like train schedules. This approach uses spatial text layout and GPT-4o to preserve essential information, ensuring accurate and error-free data interpretation.
• Guide to Speeding Up Vector Search with Minimal Accuracy Loss using Jina Embeddings to achieve 32x faster vector search performance at just a 4% cost in accuracy. It involves encoding your data as binary digits, significantly reducing storage and compute requirements.

✍️ Tutorials:

• Kate Silverstein tutorial on building local research assistant**,** set up a local, private research assistant on your laptop effortlessly with Mozilla's llamafile. Tutorial covers everything from downloading and activating the model, to connecting via LlamaIndex and managing your data.
• Plaban Nayak tutorial on Multi-document Agentic RAG using Llama-Index and Mistral.
• Diptiman Raichaudhuri tutorial on fully local Text-to-SQL using DuckDB as the database, Ollama + Mixtral-8x7B as the model, and LlamaIndex for text-to-SQL orchestration.
• Mandar Karhade tutorial on showing an end-to-end experimentation pipeline for iterating on chunk sizes, generating a synthetic dataset, and measuring how it affects evaluation metrics.

📹 Webinar:

Join us for a webinar on "Open-Source Longterm Memory for Autonomous Agents" this Thursday at 9am PT, where we'll explore the memary architecture with Julian Saks, Kevin Li, Seyeong Han and rest of memary team, diving deep into the challenges and future of long-term memory for autonomous systems.

📅 Events:

We are having our first-ever meetup at our new office in San Francisco! Join us to connect with our team and friends from Activeloop and Tryolabs, as we discuss the latest developments in generative AI.