LlamaIndex Newsletter 2023-11–07

Hi again Llama Fans! 🦙

We hope you enjoyed our OpenAI Dev Day special edition yesterday! Here’s our wrap-up of everything else that happened last week. As always, if you’ve got a project, article, or video that’s turning heads? We’re all ears! Drop us a line at news@llamaindex.ai.

And for all this goodness delivered directly to you, don’t forget to subscribe to our newsletter via our website.

🤩 First, the highlights:

1. LlamaIndex Chat: We unveiled a customizable LLM chatbot template with system prompts and avatars, all within an open-source MIT-licensed framework using LlamaIndex for TypeScript. Explore the Demo or check the Tweet.
2. Evaluator Fine-Tuning: We launched a method to enhance LLM output assessment by distilling GPT-4 into GPT-3.5, optimizing both cost and speed. See our Tweet.
3. ParamTuner: We introduced a new hyperparameter tuning abstraction to refine RAG pipeline performance, featuring objective functions, grid search, and Ray Tune integration. Check out the Notebook and Tweet.
4. CohereAI Embed v3 & Voyage AI Integration: We strengthened the LlamaIndex RAG pipeline with two powerful embedding model additions: the latest Embed v3 from CohereAI and the high-performing embedding model from Voyage AI. Tweet and tweet.

✨ Feature Releases and Enhancements:

• We introduced LlamaIndex Chat, a new feature allowing users to create and share custom LLM chatbots tailored to their data, complete with personalized system prompts and avatars. Additionally, we’re proud to share that it’s a fully open-source template under the MIT license, crafted using LlamaIndexTS for a seamless start to LLM application development. Demo, Tweet.
• We introduced a method for fine-tuning an Evaluator to distill GPT-4 into GPT-3.5, enhancing LLM output assessment while reducing costs and improving speed. Tweet.
• We introduced ParamTuner, a hyperparameter tuning abstraction for LlamaIndex RAG, streamlining the process with objective functions and support for grid search, including integration with Ray Tune for enhanced optimization. Notebook, Tweet.

🎥 Demos:

• GPTDiscord is a versatile LLM-powered Discord bot with over 20 features, including multi-modal image understanding and advanced data analysis. It boasts an infinite conversational memory and the ability to interact with various file types and internet services. Tweet.

🗺️ Guides:

• We shared a guide for integrating Activeloop’s Deep Memory with LlamaIndex, a module that enhances your embeddings at ingestion and can improve RAG metrics by 15%, all while seamlessly fitting into LlamaIndex’s automated dataset and vector store features.
• We shared a guide inspired by Chengrun Yang and GoogleDeepMind’s Optimization by Prompting paper, demonstrating how to automate prompt tuning in LlamaIndex RAG pipelines using meta-prompting, boosting evaluation performance while acknowledging the experimental nature of this technique.
• We shared a guide on how to implement Emotion Prompting in LlamaIndex, allowing you to enhance your RAG pipeline with various emotional stimuli and evaluate their impact on task performance.
• We showcased MongoDB starter kit, a comprehensive LlamaIndex RAG setup with Flask backend, Next frontend, and easy deployment to Render.

✍️ Tutorials:

• Wenqi Glantz made a blog post on deploying the HuggingFace text-embeddings-inference server on an AWS EC2 GPU instance, enhancing LlamaIndex RAG pipeline's performance and results.
• Sophia Yang’s tutorial on Zephyr-7b-beta showcases its leading capabilities in LLM technology, including how it’s benchmarked with LlamaIndex for diverse AI tasks.
• Sudarshan Koirala gave a tutorial on how to build a multi-modal retrieval system with LlamaIndex, Qdrant, and bge/CLIP embeddings.
• Sophia Yang’s gave another tutorial, this time on Small-to-Big Retrieval with LlamaIndex in building advanced RAG systems.
• Ravi Theja’s tutorial on the Router Query Engine that helps you to set up multiple indices/ query engines for your dataset, allowing the LLM to choose the most suitable one for each specific question.

⚙️ Integrations & Collaborations:

• We integrated the Tavily AI research API into the LlamaIndex RAG pipeline, offering a robust tool for web research to enhance LLM agent automation. Notebook, Tweet.
• We integrated Noam Gat’s LLM Enforcer into the LlamaIndex RAG pipeline to ensure structured outputs for various models. Docs, Tweet.
• We integrated the latest Embed v3 model from CohereAI, enhancing document retrieval quality within the LlamaIndex RAG pipeline. Notebook, Tweet.
• We integrated the new Voyage AI embedding model, a top-performing option for RAG pipelines. Notebook, Tweet.