TWIL – Week of January 27, 2025

Table of Contents

This week, I delved into advancements in AI privacy and tools, notably Private AI’s solutions for data anonymization. I explored innovative AI tools like Browser Use’s web-ui for browser-accessible LLMs, the screenshot-to-code tool, and pandas-ai for engaging with dataframes conversationally.

In the LLM realm, I discovered Anthropic’s support for RAG citations, Ollama’s local LLM serving, and frameworks like AutoMetaRAG and MAIN-RAG for enhancing data retrieval. I also learned about Amazon Bedrock’s metadata filtering enhancements and Azure AI Search’s new capabilities, including query rewriting and document summarization. Lastly, I looked into uv, a promising new Python package manager highlighted for its speed and efficiency.

AI Privacy

AI Tools

LLMs

  • Anthropic supports citations for RAG out of the box!
  • Ollama is a service that allows easy and quick serving of LLMs locally.
  • AutoMetaRAG gives an LLM a description of the data and probable questions and the LLM gives a proposed metadata to use in the vector store. The metadata options are given to an inference LLM at runtime, which will select metadata for pre-filtering retrieval.
  • MAIN-RAG is a framework for using a multi-agent system to filter out irrelevant documents and appropriately sort relevant ones from a RAG retriever.
  • Amazon Bedrock implements intelligent metadata filtering using Pydantic data models and function calling.

Vector Databases

Python

  • uv is an extremely fast Python package/project manager. ML Engineer Larry Du wrote in June 2024 that uv was the most promising up-and-coming python package managers.