AI Bricks

AI Bricks provide minimalistic toolbox for creating Generative AI based systems.

Features:

• unified interface to multiple Generative AI providers
• strong support for local model servers (KoboldCpp, LMStudio, Ollama, LlamaCpp, tabbyAPI, ...)
• ability to record requests and responses (in sqlite) and generate reports (ie usage)
• configuration driven approach to prompt templates (yaml + jinja2)
• minimal dependencies (requests, yamja)

AI Bricks focuses on providing basic building blocks rather than a complex framework. This aligns with research showing that the most successful LLM implementations use simple, composable patterns rather than complex frameworks. By keeping things minimal, it allows developers to build exactly what they need without unnecessary abstractions.

Examples

Example from the aisuite repo:

import aibricks
client = aibricks.client()

models = ["openai:gpt-4o", "xai:grok-beta"]

messages = [
    {"role": "system", "content": "Respond in Pirate English."},
    {"role": "user", "content": "Tell me a joke."},
]

for model in models:
    response = client.chat(
        model=model,
        messages=messages,
        temperature=0.75
    )
    print(response['choices'][0]['message']['content'])

Minimalistic implementation of MemGPT-like agent

import aibricks
from pathlib import Path

client = aibricks.client(model='xai:grok-beta')
config = aibricks.load_config(Path(__file__).with_name('memgpt_v1.yaml'))

memory = ''
messages = [{'role': 'system', 'content': config.render('system_prompt', memory=memory)}]
N_MESSAGES_TO_KEEP = 4


def chat_turn(user_input):
    global messages, memory
    # get model response
    messages += [{'role': 'user', 'content': user_input}]
    response = client.chat(messages=messages)
    content = response['choices'][0]['message']['content']
    messages += [{'role': 'assistant', 'content': content}]
    # handle tool usage
    for tag, attr, kw in aibricks.parse_xml(content):
        tool = attr['tool']
        if tool == 'respond':
            print('AI:', kw['message'])
        elif tool == 'memory.add':
            memory += kw['new'] + '\n'
        elif tool == 'memory.replace':
            memory = memory.replace(kw['old'], kw['new'])
    # update system prompt with new memory
    messages[0]['content'] = config.render('system_prompt', memory=memory)
    # evict old messages
    messages = messages[0:1] + messages[1:][-N_MESSAGES_TO_KEEP:]


if __name__ == '__main__':
    chat_turn("Hi! My name is Maciej and I'm from Poland!")
    chat_turn("I was born in 1981 and learned to program when I was 8.")
    chat_turn("Can you guess my first computer and programming language?")
    chat_turn("Well, it was Timex 2048 and BASIC.")
    print('\nMEMORY:\n' + memory)

This example can be found in the examples/memgpt directory.

Supported providers

Provider	Example Connection String	Environmental Variables	Notes
OpenAI	openai:gpt-4o-mini	OPENAI_API_KEY
Anthropic	anthropic:claude-3-5-haiku-latest	ANTHROPIC_API_KEY
XAI	xai:grok-beta	XAI_API_KEY
Google	google:gemini-1.5-flash	GEMINI_API_KEY
OpenRouter	openrouter:openai/gpt-4o	OPENROUTER_API_KEY
ArliAI	arliai:Llama-3.1-70B-Tulu-2	ARLIAI_API_KEY
Together	together:google/gemma-2b-it	TOGETHER_API_KEY
Ollama	ollama:qwen2.5-coder:7b	-	GGUF
LMStudio	lmstudio:qwen2.5-14b-instruct	-	GGUF dynamic model loading
KoboldCpp	koboldcpp	-	GGUF
LlamaCpp	llamacpp	-	GGUF
tabbyAPI	tabbyapi	TABBYAPI_API_KEY	EXL2, GPTQ
dummy	dummy	-

License

MIT

Installation

Don't. The project is still in its infancy.

References

• Building Effective Agents
• Building an AI-Powered Game
• LLMs as Operating Systems: Agent Memory
• Letta (formerly MemGPT)
• MemGPT: Towards LLMs as Operating Systems
• OpenAI Chat API
• aisuite repo
• ai-bricks-v3 repo