oneping 0.5.5

LLM provider abstraction layer.

oneping

oneping.reply('Give me a ping, Vasily. One ping only, please.', provider='anthropic')

This is a Python library for querying LLM providers such as OpenAI or Anthropic, as well as local models. The main goal is to create an abstraction layer that makes switching between them seamless. Currently the following providers are supported: openai, anthropic, fireworks, and local (local models).

There is also a Chat interface that automatically tracks the message history. Kind of departing from the "one ping" notion, but oh well. Additionally, there is a textual powered console interface and a fasthtml powered web interface. Both are components that can be embedded in other applications.

Requesting the local provider will target localhost and use an OpenAI-compatible API as in llama.cpp or llama-cpp-python. The various native libraries are soft dependencies and the library can still partially function with or without any or all of them. The native packages for these providers are: openai, anthropic, and fireworks-ai.

Installation

For standard usage, install with:

pip install oneping

To install the native provider dependencies add "[native]" after oneping in the command above. The same goes for the chat interface dependencies with "[chat]".

The easiest way to handle authentication is to set an API key environment variable such as: OPENAI_API_KEY, ANTHROPIC_API_KEY, FIREWORKS_API_KEY, etc. You can also pass the api_key argument to any of the functions directly.

Library Usage

Basic usage with Anthropic through the URL interface:

response = oneping.reply(query, provider='anthropic')

The reply function accepts a number of arguments including (some of these have per-provider defaults):

• query (required): The query to send to the LLM (required)
• provider = local: The provider to use: openai, anthropic, fireworks, or local
• system = None: The system prompt to use (not required, but recommended)
• prefill = None: Start "assistant" response with a string (Anthropic doesn't like newlines in this)
• model = None: Indicate the desired model for the provider (provider default)
• max_tokens = 1024: The maximum number of tokens to return
• history = None: List of prior messages or True to request full history as return value
• native = False: Use the native provider libraries
• url = None: Override the default URL for the provider (provider default)
• port = 8000: Which port to use for local or custom provider
• api_key = None: The API key to use for non-local providers

For example, to use the OpenAI API with a custom system prompt:

response = oneping.reply(query, provider='openai', system=system)

To conduct a full conversation with a local LLM, see Chat interface below. For streaming, use the function stream and for async streaming, use stream_async. Both of these take the same arguments as reply.

Command Line

You can call oneping directly or as a module with python -m oneping and use the following subcommands:

• reply: get a single response from the LLM
• stream: stream a response from the LLM
• embed: get embeddings from the LLM
• console: start a console (Textual) chat
• web: start a web (FastHTML) chat

These accept the arguments listed above for reply as command line arguments. For example:

oneping stream "Does Jupiter have a solid core?" --provider anthropic

Or you can pipe in your query from stdin:

echo "Does Jupiter have a solid core?" | oneping stream --provider anthropic

I've personally found it useful to set up aliases like claude = oneping stream --provider anthropic.

Chat Interface

The Chat interface is a simple wrapper for a conversation history. It can be used to chat with an LLM provider or to simply maintain a conversation history for your bot. If takes the usual reply, stream, and stream_async functions, and calling it directly will map to reply.

chat = oneping.Chat(provider='anthropic', system=system)
reply1 = chat(query1)
reply2 = chat(query2)

There is also a textual powered console interface and a fasthtml powered web interface. You can call these with: oneping console or oneping web.

Server

The server module includes a simple function to start a llama-cpp-python server on the fly (oneping.server.start in Python or oneping server from the command line).

oneping server <path-to-gguf>

To run the server in embedding mode, pass the --embedding flag. You can also specify things like --host and --port or any options supported by llama-cpp-python.

Embeddings

Embeddings queries are supported through the embed function. It accepts the relevant arguments from the reply function. Right now only openai and local providers are supported.

vecs = oneping.embed(text, provider='openai')

and on the command line:

oneping embed "hello world" --provider openai