LLM

The LLM extension is still experimental.

Stability is not yet guaranteed and usage may still change.

The LLM extension enables integration with Large Language Models (LLMs) in two ways:

• An model context protocol (MCP) server allows them to access medical logic modules as tools
• Plugins allow MLMs to prompt LLMs through READ statements. Requests are sent to a configured OpenAI-compatible endpoint.

In addition, the server provides an OpenAI chat completion request proxy that automatically adds tool LLMs into the request before forwarding it to the configured provider.

MCP Server

INFO

Only MLMs that are explicitly configured in the tool list will be available to LLMs.

On the LLM extension tab, enable MCP with the toggle switch and add eligible MLMs through the + button. You can click the X on added MLMs to remove them again.

Eligibility

An MLM is only eligible for MCP if it fulfils two conditions:

• It must have content in the purpose slot
• It must have a special input argument definition as explanation slot:

  explanation: 
    Argument <name> of type <type> described <description>;
    Argument <name> of type <type> described <description>;
    ...
  ;;

Support for input and output definitions is planned in a future Arden Syntax version, and the extension may be revised to use that definition when it is published.

Plugins

Plugins allow MLMs to prompt LLMs through READ statements.

The parameters to be provided are:

• Base URL: An OpenAI-compatible endpoint to use
• Model: The specific model identifier included in the model field of OpenAI requests
• API Key: API included in the Authorization Bearer header, usually the key or token configured with your provider. Put in a dummy value like EMPTY for authentication-free endpoints.
• System Prompt: An optional predefined instruction that sets behavior, tone, or constraints for the LLM across all interactions

Usage

Provide at least a prompt to a READ statement. You can optionally override the model too.

data:
    LET response be READ { FROM my-llm-plugin WITH PROMPT "Say hi" };
    LET response be READ { FROM my-llm-plugin WITH PROMPT "Say hi" AND MODEL "mistral:7b" };

Returned is a string of the generated response.

Language models prompted this way can themselves choose to call other MLMs as tools, opening up the possibility for interesting reasoning chains.

Testing the setup

A chat button is available on the right side of each configuration to validate your configuration and play with the capabilities. This testing interface allows to submit prompts directly to the configured LLM and includes the configured tools.

If the model responds correctly and tool interactions behave as expected, configuration is validated.