LMML (Language Model Markup Language) is a proposed and open standard for giving agential configuration to a predictive API call on any large language model. As it is an open standard, we invite anyone to join, fork, and contribute to this repository. Join Our Discord to discuss.
The LMML Engine acts as an interpreter between a standardized document format for autonomous agent configuration and the final text prompt fed to the API endpoint of a Large Language Model. In other words, as the world of AI assisted applications builds up, Prompting is Programming; meaning that in order to get the results one wants directly from an LLM (Large Language Model), one has to know exactly how to warm up the model with system prompts, and coerce the model to output what one wants.
In short, the niche that LMML Engine fits is to Agentialize and intialize AI prompt streams (conversations with different chat agents) that can be defined by an LMML document sitting at a standard internet URI. The LMML Engine can power a new type of chat based browser that simply loads in LMML documents and then uses the user's own API credentials from the user device to perform API calls, or uses the App developer credentials from the server.
I am of the opinion that HTML is a language which is easy to learn, concise, and serves a purpose of showing a web page. Prompting, being the future of programming is going to need some standards that put this explosion of prompt types and state-of-the-art into an easy to remember and reference format which has all of the caveats of a web document. The future of AI and AGI based agents online will be an open standard, and we might as well stick closely to one we already know.
It looks a lot like HTML except it is tailored toward the types of things people want to put into a prompt for a large language model such as GPT, Bloom, LLamA, AlPaca, including information about which model this task is built for.
<LMML version="1.0">
<AGENT api="openai" model="gpt-3.5-turbo" name="Researcher" id="researcher">
<SELF>
<EXTEND src="https://some.site/some/agent.lmml"></EXTEND>
<ABOUT>You are a researcher who excels at generating search terms and finding lists of possible leads to solve a given problem or provide research reports on a recent topic. Leads are delivered in the form of a list of the most likely URLs. You can also find leads in the form of contacts. You self improve through self reflection and learning from mistakes.</ABOUT>
</SELF>
<MEMORY>
<TASK id="search" name="Search">
<MESSAGE id="">Search for "<INPUT name="" label=""/>"</MESSAGE>
<OUTPUT>
<IMPERATIVE name="reflect" >REFLECT</IMPERATIVE>
<IMPERATIVE name="mostlikely">MOST LIKELY</IMPERATIVE>
</OUTPUT>
</TASK>
<TASK id="get_contact_details" src="./contact_finder.lmml"/>
</MEMORY>
</AGENT>
</LMML>This standard can streamline the act of creating autonomous agents using large language model APIs in a way that will easily interconnect these agents online.
Imagine an experience where you open a chat app and this is your window to the entire internet. The agent you speak with is one you are familiar with and remembers your previous conversations. The agent can also talk with other agents online and you can view their conversation right in the chat history. The browser on your device has any API keys already added into it for any AI service you might want to use specifically. You can go directly to chat with some AI chat agent on a website by simply navigating to that website.