llmuri -- URI support for LLMs

Standard naming support for LLMs: llm:myhost/mymodel?params...

This project defines a standard URI for specifying LLMs and provides functionality to convert LLM strings for use with popular LLM packages.

Example

from llmuri import uri_to_litellm
from litellm import completion

for uri in [
    "llm:openai/gpt-4",
    "openai/gpt-4",
    "llm:ollama/llama2",
    "llm:mistralai/mistral-medium?temperature=0.2&max_tokens=4",
    "llm:ollama@example.com:11434/llama2",
]:
    kwargs = uri_to_litellm(uri)
    response = completion(
        messages=[{"content": "respond in 20 words. who are you?", "role": "user"}],
        **kwargs,
    )
    print(response)

Installation

pip install llmuri

URI Specification

[scheme:]provider[@host[:port]]/model-name[?parameter1=value1[&parameter2=value2]...

examples:

ollama/llama2
llm:ollama/llama2?temperature=0.2
llms:ollama@example.com:11434/llama2

• scheme can be "llm" or "llms" to specify that the LLM is hosted behind an https web service. If the scheme is ommitted, "llm" is assumed.

• provider is the name of the LLM API or service.

• host is an optional hostname where the LLM service is located.

• port is the service's port number.

• model-name is the name of the model to be executed.

• parameter-list is a list of model-specific parameters.

Canonical Abbreviations

The canonical set of well-known services (e.g. "openai", "mistralai") is in canonical-abbreviations.csv.

In-process References

If the port portion of the URI is "mem", then the LLM is not accessed externally, but refers to an LLM instantiated in the current process.

The semantics of in-process references is a WIP and subject to change.

• provider and host are available for llm specification.

llm:A@B:mem/mymodel  #  A,B are available
llmp:A@B:C/mymodel   #  A,B,C are available if we introduce a new scheme. bad idea?

Examples:

llm:MyTransformer@:mem        # Instantiate an instance of MyTransformer
llm:openai@:mem/gpt-3.5-turbo # Instantiate an instance of OpenAI() using gpt-3.5-turbo
llm:completions@mistralai:mem/mistral-medium  # completions interface?