TypeAgent explores an architecture for building a single personal agent with natural language interfaces, leveraging LLM advancements. It targets developers and researchers interested in safely combining stochastic LLMs with traditional software components for task automation. The primary benefit is enabling a unified agent to interact with various applications through natural language.

How It Works

TypeAgent is built on three core principles: distilling models into logical structures, using structure to control information density, and using structure to enable collaboration. It employs a system called AMP (Actions, Memories, and Plans) to integrate these elements. Actions are defined using TypeChat schemas, and memory is managed via Structured RAG, an indexing and query processing approach designed for enhanced conversational recall and precision compared to Classic RAG.

Quick Start & Requirements

• Install/Run: Follow step-by-step instructions for Windows WSL2, Linux (Ubuntu/Debian), or MacOS. Detailed setup is in the TypeScript code directory's README.
• Prerequisites: Requires Azure OpenAI services (API keys needed), and is tested in English.
• Resources: Setup involves provisioning endpoints and potentially local state storage.
• Links: TypeAgent Shell Example, TypeScript Code README

Highlighted Details

• Structured RAG for agent memory offers improved recall for conversational context.
• TypeChat schemas are used to define agent actions and validate LLM responses.
• TypeAgent Cache aims to reduce LLM costs and latency by caching action translations.
• Extensible architecture allows for custom agents to be plugged into the TypeAgent Shell.

Maintenance & Community

• This is a Microsoft project with an active development cycle and frequent updates.
• Questions can be posted in the Q&A discussion section.
• Follows the Microsoft Open Source Code of Conduct.

Licensing & Compatibility

• The repository contains code under various licenses, including MIT and Apache 2.0. Specific license details for each component should be verified.
• Sample agents are not intended for real-world use without further testing and validation.

Limitations & Caveats

TypeAgent is early-stage sample code and not a framework; all code is for building examples. It is actively developed with frequent refactoring. Performance may vary in languages other than English. Agent validity depends on the accuracy of its schema representation and LLM responses.