Skip to main content

Plugins Technical Details

OpenAssistant incorporates a basic implementation of support for external augmentation using AI plugins of the OpenAI spec. In the current state, this is more of a proof-of-concept and should be considered experimental.

Architecture

There is now middleware between work.py (in the worker) and the final prompt that is passed to the inference model server for generation. That middleware is responsible for checking if there is an enabled plugin in the UI. If so, it will add curated pre-prompts for plugin usage and generate subsequent calls to the LLM (inner monologues) to generate the final externally augmented prompt, that will be passed back to the inference model server, for final LLM generation and streaming of tokens to the frontend.

Plugins

Plugins are in essence just wrappers around APIs to help LLM utilize an API more precisely and reliably. They can be quite useful and powerful augmentation tools for OpenAssistant. The two main parts of a plugin are the ai-plugin.json file, which is just the main descriptor of a plugin, and the OpenAPI specification of the plugin APIs.

Here is the OpenAI plugin specification that is currently partially supported by OpenAssistant. Instructions for creating and hosting a plugin as well as a template for basic plugins can be found here.

For now, only non-authentication-based and only (GET request) plugins are supported. Some of them are:

And quite a few of them can be found on the plugin "store" at wellknown.ai

Adding support for all other request types is tricky with the current approach. It would be best to drop current prompts explaining in-depth the API to the LLM and instead just show it complete JSON/YAML content. But for that to work as well as the current approach we would need larger context size and more capable models, which are still in development.

We provide a few official OA plugins, and any community member can develop a plugin.

Notes on reliability, performance, and limitations of the system

  • Performance can vary a lot depending on the models and plugins used.

  • Performance and consistency should improve as we get better and better models. One of the biggest limitations at the moment is context size and instruction following capabilities.

  • We combat model limitations with prompt tricks, truncations of plugin OpenAPI descriptions and dynamically including/excluding parts of prompts in internal processing of subsequent generations of intermediate texts (inner monologues).

  • The current approach is somewhat hybrid and relies on the zero-shot capabilities of a model.

Relevant files for the inference side of the plugin system

  • chat_chain.py
  • chatchain_utils.py (tweaking tools/plugin description string generation can help for some models)_
  • chatchain_prompts.py (tweaking prompts can help also)_