🖼️ Infra ✨ Feature? ❤️🩹 Bug? 📞 Help?
- Code Interpreter: Generate and runs Python code in a sandboxed environment autonomously. (beta)
- Knowledge Retrieval: Retrieves external knowledge or documents autonomously.
- Function Calling: Defines and executes custom functions autonomously.
- Actions: Execute requests to external APIs autonomously.
- Files: Supports a range of file formats.
- OpenAI compatible: Works with OpenAI (Assistants) SDK
![hal-9100-2](https://private-user-images.githubusercontent.com/25003283/322167835-5a393d61-7a1d-4e06-8932-f822b18015ba.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.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.saTaTxrd1OQdsleq7dBMmBJbEWCdx_LldysY5shDueo)
- You want to increase customization (e.g. use your own models, extend the API, etc.)
- You work in a data-sensitive environment (healthcare, IoT, military, law, etc.)
- Your product does have poor or no internet access (military, IoT, edge, extreme environment, etc.)
- (not our main focus) You operate on a large scale and want to reduce your costs
- (not our main focus) You operate on a large scale and want to increase your speed
First, our definition of Software 3.0, as it is a loaded term: Software 3.0 is the bridge connecting the cognitive capabilities of Large Language Models with the practical needs of human digital activity. It is a comprehensive approach that allows LLMs to:
- perform the same activity (or better) on the digital world than humans
- generally, allow the user to perform more operations without conscious effort
HAL-9100 is in continuous development, with the aim of always offering better infrastructure for Edge Software 3.0. To achieve this, it is based on several principles that define its functionality and scope.
Less prompt is more
As few prompts as possible should be hard-coded into the infrastructure, just enough to bridge the gap between Software 1.0 and Software 3.0 and give the client as much control as possible on the prompts.
Edge-first
HAL-9100 does not require internet access by focusing on open source LLMs. Which means you own your data and your models. It runs on a Raspberry PI (LLM included).
OpenAI-compatible
OpenAI spent a large amount of the best brain power to design this API, which makes it an incredible experience for developers. Support for OpenAI LLMs are not a priority at all though.
Reliable and deterministic
HAL-9100 focus on reliability and being as deterministic as possible by default. That's why everything has to be tested and benchmarked.
Flexible
A minimal number of hard-coded prompts and behaviors, a wide range of models, infrastructure components and deployment options and it play well with the open-source ecosystem, while only integrating projects that have stood the test of time.
Get started in less than a minute through GitHub Codespaces:
Or:
git clone https://github.com/llm-edge/hal-9100
cd hal-9100
To get started quickly, let's use Anyscale API.
Get an API key from Anyscale. You can get it here. Replace in hal-9100.toml the model_api_key
with your API key.
Usage w/ ollama
- use
model_url = "http://localhost:11434/v1/chat/completions"
- set
gemma:2b
in examples/quickstart.js - and run
ollama run gemma:2b & && docker compose --profile api -f docker/docker-compose.yml up
Install OpenAI SDK: npm i openai
Start the infra:
docker compose --profile api -f docker/docker-compose.yml up
Run the quickstart:
node examples/quickstart.js
Is there a hosted version?
No. HAL-9100 is not a hosted service. It's a software that you can deploy on your infrastructure. We can help you deploy it on your infrastructure. Contact us.
Which LLM API can I use?
Examples of LLM APIs that does support OpenAI API-like, that you can use:
- ollama
- MLC-LLM
- FastChat (good if you have a mac)
- vLLM (good if you have a modern gpu)
- Perplexity API
- Mistral API
- anyscale
- together ai
We recommend these models:
- mistralai/Mixtral-8x7B-Instruct-v0.1
- mistralai/mistral-7b
Other models have not been extensively tested and may not work as expected, but you can try them.
What's the difference with LangChain?
1. LangChain spans proprietary LLM and open source, among the thousands of things it spans. HAL-9100 laser focuses on Software 3.0 for the edge.- You can write AI products in 50 lines of code instead of 5000 and having to learn a whole new abstraction