Works with most commercial/remote and open/local LLMs.

Set up Multi-agent, multi-LLM system: use stronger LLMs for agents requiring strong reasoning and instruction-following, and delegate simpler tasks to weaker/local LLMs.

Supports OpenAI function-calling as well as native equivalent called ToolMessage, which works with LLMs thatdo not have built-in function-calling. Simply specify structure as a (nested) Pydantic object.

Batteries-included: vector-databases for RAG (Retrieval-Augmented Generation), caching, logging/observability.

Specialized agents available: DocChatAgent, SQLChatAgent, TableChatAgent (for tabular data, e.g. csv/dataframes).

DocChatAgent handles text, PDF, Docx files/URLS, and has state-of-the art techniquesfor retrieval combining lexical and semantic search.

Documentation: https://langroid.github.io/langroid/