Repository files navigation

TypeAgent issample code that explores an architecture for building asingle personal agent withnatural language interfaces leveraging current advances in LLM technology.

The goal of the TypeAgent team is to explore how to get work done by safely and efficiently combining stochastic systems like language models with traditional software components.  Three principles have emerged during this investigation.  They are listed below along with examples of how the principles apply to actions, memory and plans.

• Principle: distill models into logical structures
  • Actions: find translation patterns and replace some model calls by applying patterns
  • Memory: build ontologies from text
  • Plans: people, programs and models collaborate using “tree of thought”
• Principle: use structure to control information density
  • Actions: applications define discrete categories with dense descriptions of action sets
  • Memory: tight semantic structures fit into attention budget
  • Plans: each search tree node defines a focused sub-problem
• Principle: use structure to enable collaboration
  • Actions: humans decide how to disambiguate action requests
  • Memory: simple models extract logical structure from text
  • Plan: quality models, advantage models, language models, humans and programs collaborate to expand each best-first-search node

We are trying to create human-like memory with super-human precision and recall for agent conversations.  We are using a new indexing and query processing approach calledStructured RAG as the basis for agent memory.  Structured RAG does substantially better than Classic RAG at answering questions about past conversations such as "what were the books we talked about?" and "what step were we on in building the photo montage?"

We are trying to build a single personal agent that can apply to any application.  To apply agent interfaces to all applications, we need to map user requests to actions at much lower cost and latency than current systems.  To make this possible, we have created a system that can distill language models into logical systems that can handle most user requests.

Actions and memories flow together.  Actions like "add to my calendar pickle ball game 2-3pm on Friday" yield memories that can become parameters of future actions like "put in an hour of recovery time after my pickle ball game."  We are working on an architecture, AMP, that enables this natural information flow by integrating actions, memories, and plans

We are applying AMP to the web by creating a browser that enables web sites to register actions through a JavaScript interface.

TypeAgent Shell example is the starting point to explore thesingle personal agent withnatural language interfaces we have built so far. It is an Electron application for interacting with multiple registered agents using a single unified user interface. TypeAgent Shell includes:

• Single personal agent conversational interface with voice support
• Collaborate with users to perform and dispatch actions to an extensible set of agents, answer question and carry on a conversation.
• Conversational memory based on Structured RAG
• Integration with TypeAgent Cache to lower cost and latency

Follow these step-by-step instructions to quickly setup tools and environments from scratch to build, run, explore, and develop.

• Windows
• WSL2
• Linux (Ubuntu/Debian)
• MacOS

For more detailed setup instructions, see theREADME.md in the TypeScript codedirectory

• TypeAgent Dispatcher

  Explores applying structured prompting and LLM to route user requests to agents whose typed contract best matches user intent. Main component of the personal agent.

• KnowPro

  Explores how to implement agent memory using the ideas ofStructured RAG.

• TypeAgent Cache

  Explores how LLM with structured prompting can be used to cache action translation, minimizing the need to go the LLM.

Storage, registration, chat, memory and other state maintained by examples istypically storedlocally inyour user folder on your development machine. State is typically saved as ordinary text or JSON files in sub-folders below your user folder.

Example agents that use the Microsoft Graph or similar external services may store state in those services.

Code in this repo doesn't not collect telemetry by default.

• TypeAgent is sample code shared to encourage the exploration of natural language agent architectures using structured prompting and LLM
• Sample agents are not intended to be implemented in real-world settings without further testing/validation.

• Publish libraries for agent memory and action dispatch.

TypeAgent is early stage sample code. TypeAgent is not a framework. All code in this repo is intended for building examples (apps, agents, and dispatcher hosts) only.

• TypeAgent is inactive development with frequent updates and refactoring.
• TypeAgent has been tested with Azure Open AI services on developer's own machines only.
• TypeAgent is currently tested in English. Performance may vary in other languages.
• TypeAgent uses schema to validate LLM responses. An agent's validity therefore depends on how wellits schema represents the user intents and LLM responsesfor its domains.
• You are responsible for supplying anyAPI keys for services used by examples. You can check theAzure provisioning readme for a quickstart on setting up the necessary endpoints if you do not already have endpoints.

This repo contains the personal agent and example application agents, along with internal packages used to build them.TypeAgent uses structured prompting with LLM technique for many of the components,to build a set of application agents thattake actions. Agents define actions usingTypeChat schemas.

TypeAgent Shell's functionality can be extended by installing/registering additional agents. For developers who are interested in experimenting adding action dispatch for their own scenarios, they can create acustom agents that plugs into theTypeAgent Shell example to explore using thedispatcher to route actions to their custom agents. TheEcho agenttutorial is a starting point to create a plugin agent, andTypeAgent SDK provides the interface definitions betweendispatcher and the agent.

For developers who want to modify TypeAgent or contribute to our repo.

Microsoft TypeAgent Repo is a mono-repo, with components organized with the following root folders based on language used.

• ts TypeScript code (Readme)
• python Python code (Readme)
• dotnet Dotnet (C#) code (Readme)
• android Android (Kotlin/Java) code (Readme)

See more information about working with the repohere.

• TypeAgent Shell
• TypeAgent CLI

• Application agents with natural language interfaces integrated withTypeAgent Shell andTypeAgent CLI

  • Music Player
  • Chat
  • Browser
  • VS Code
  • List Management
  • Calendar
  • Email
  • Desktop
  • Image
  • Markdown
  • Montage
  • Spelunker
  • Turtle
  • Phone
  • Photo
  • androidMobile

If you have any questions about our project, you can post them in ourQ&A discussion section.

This project has adopted theMicrosoft Open Source Code of Conduct.For more information see theCode of Conduct FAQ orcontactopencode@microsoft.com with any additional questions or comments.

This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsofttrademarks or logos is subject to and must followMicrosoft's Trademark & Brand Guidelines.Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship.Any use of third-party trademarks or logos are subject to those third-party's policies.