AutoGPT, babyAGI, ... and now LocalAGI!
LocalAGI is a small 🤖 virtual assistant that you can run locally, made by the LocalAI author and powered by it.
The goal is:
- Keep it simple, hackable and easy to understand
- No API keys needed, No cloud services needed, 100% Local. Tailored for Local use, however still compatible with OpenAI.
- Smart-agent/virtual assistant that can do tasks
- Small set of dependencies
- Run with Docker/Podman/Containers
- Rather than trying to do everything, provide a good starting point for other projects
Note: Be warned! It was hacked in a weekend, and it's just an experiment to see what can be done with local LLMs.
- 🧠 LLM for intent detection
- 🧠 Uses functions for actions
- 📝 Write to long-term memory
- 📖 Read from long-term memory
- 🌐 Internet access for search
- 🗃️ Write files
- 🔌 Plan steps to achieve a goal
- 🤖 Avatar creation with Stable Diffusion
- 🗨️ Conversational
- 🗣️ Voice synthesis with TTS
Search on internet (interactive mode)
search.mp4
Plan a road trip (batch mode)
planner.mp4
Note: The demo is with a GPU and
30b
models size
No frills, just run docker-compose and start chatting with your virtual assistant:
# Modify the configuration
# vim .env
# first run (and pulling the container)
docker-compose up
# next runs
docker-compose run -i --rm localagi
By default localagi starts in interactive mode
Road trip planner by limiting searching to internet to 3 results only:
docker-compose run -i --rm localagi \
--skip-avatar \
--subtask-context \
--postprocess \
--search-results 3 \
--prompt "prepare a plan for my roadtrip to san francisco"
Limit results of planning to 3 steps:
docker-compose run -i --rm localagi \
--skip-avatar \
--subtask-context \
--postprocess \
--search-results 1 \
--prompt "do a plan for my roadtrip to san francisco" \
--plan-message "The assistant replies with a plan of 3 steps to answer the request with a list of subtasks with logical steps. The reasoning includes a self-contained, detailed and descriptive instruction to fullfill the task."
localagi has several options in the CLI to tweak the experience:
--system-prompt
is the system prompt to use. If not specified, it will use none.--prompt
is the prompt to use for batch mode. If not specified, it will default to interactive mode.--interactive
is the interactive mode. When used with--prompt
will drop you in an interactive session after the first prompt is evaluated.--skip-avatar
will skip avatar creation. Useful if you want to run it in a headless environment.--re-evaluate
will re-evaluate if another action is needed or we have completed the user request.--postprocess
will postprocess the reasoning for analysis.--subtask-context
will include context in subtasks.--search-results
is the number of search results to use.--plan-message
is the message to use during planning. You can override the message for example to force a plan to have a different message.--tts-api-base
is the TTS API base. Defaults tohttps://api:8080
.--localai-api-base
is the LocalAI API base. Defaults tohttps://api:8080
.--images-api-base
is the Images API base. Defaults tohttps://api:8080
.--embeddings-api-base
is the Embeddings API base. Defaults tohttps://api:8080
.--functions-model
is the functions model to use. Defaults tofunctions
.--embeddings-model
is the embeddings model to use. Defaults toall-MiniLM-L6-v2
.--llm-model
is the LLM model to use. Defaults togpt-4
.--tts-model
is the TTS model to use. Defaults toen-us-kathleen-low.onnx
.--stablediffusion-model
is the Stable Diffusion model to use. Defaults tostablediffusion
.--stablediffusion-prompt
is the Stable Diffusion prompt to use. Defaults toDEFAULT_PROMPT
.--force-action
will force a specific action.--debug
will enable debug mode.
To use a different model, you can see the examples in the config
folder.
To select a model, modify the .env
file and change the PRELOAD_MODELS_CONFIG
variable to use a different configuration file.
The "goodness" of a model has a big impact on how LocalAGI works. Currently 13b
models are powerful enough to actually able to perform multi-step tasks or do more actions. However, it is quite slow when running on CPU (no big surprise here).
The context size is a limitation - you can find in the config
examples to run with superhot 8k context size, but the quality is not good enough to perform complex tasks.
It is a dead simple experiment to show how to tie the various LocalAI functionalities to create a virtual assistant that can do tasks. It is simple on purpose, trying to be minimalistic and easy to understand and customize for everyone.
It is different from babyAGI or AutoGPT as it uses LocalAI functions - it is a from scratch attempt built on purpose to run locally with LocalAI (no API keys needed!) instead of expensive, cloud services. It sets apart from other projects as it strives to be small, and easy to fork on.
LocalAGI
just does the minimal around LocalAI functions to create a virtual assistant that can do generic tasks. It works by an endless loop of intent detection
, function invocation
, self-evaluation
and reply generation
(if it decides to reply! :)). The agent is capable of planning complex tasks by invoking multiple functions, and remember things from the conversation.
In a nutshell, it goes like this:
- Decide based on the conversation history if it needs to take an action by using functions. It uses the LLM to detect the intent from the conversation.
- if it need to take an action (e.g. "remember something from the conversation" ) or generate complex tasks ( executing a chain of functions to achieve a goal ) it invokes the functions
- it re-evaluates if it needs to do any other action
- return the result back to the LLM to generate a reply for the user
Under the hood LocalAI converts functions to llama.cpp BNF grammars. While OpenAI fine-tuned a model to reply to functions, LocalAI constrains the LLM to follow grammars. This is a much more efficient way to do it, and it is also more flexible as you can define your own functions and grammars. For learning more about this, check out the LocalAI documentation and my tweet that explains how it works under the hoods: https://twitter.com/mudler_it/status/1675524071457533953.
The intention of this project is to keep the agent minimal, so can be built on top of it or forked. The agent is capable of doing the following functions:
- remember something from the conversation
- recall something from the conversation
- search something from the internet
- plan a complex task by invoking multiple functions
- write files to disk
- 100% Local, with Local AI. NO API KEYS NEEDED!
- Create a simple virtual assistant
- Make the virtual assistant do functions like store long-term memory and autonomously search between them when needed
- Create the assistant avatar with Stable Diffusion
- Give it a voice
- Use weaviate instead of Chroma
- Get voice input (push to talk or wakeword)
- Make a REST API (OpenAI compliant?) so can be plugged by e.g. a third party service
- Take a system prompt so can act with a "character" (e.g. "answer in rick and morty style")
Run docker-compose with main.py checked-out:
docker-compose run -v main.py:/app/main.py -i --rm localagi
- a 13b model is enough for doing contextualized research and search/retrieve memory
- a 30b model is enough to generate a roadmap trip plan ( so cool! )
- With superhot models looses its magic, but maybe suitable for search
- Context size is your enemy.
--postprocess
some times helps, but not always - It can be silly!
- It is slow on CPU, don't expect
7b
models to perform good, and13b
models perform better but on CPU are quite slow.