Cool project!

[sombraSoft]

Hey @orthagonal , I found your repo when searching for a Langchain implementation for elixir.
Are you accepting PRs? Let me know what are your plans for the repo!

[orthagonal]

Thanks a bunch! Yes I am accepting PRs, suggestions, advice, personal insults, etc. My tentative roadmap for first phase is:

• basic docs for all the modules and functions (I'm doing that now).

• "anchors", a @behavior at the end of a Chain where it does something to get in alignment, usually asking a human being or hard-coded program for confirmation before proceeding. The default implementation is for IO.puts/gets command line but the idea is you can implement your own behaviors for web, voice, gesture, whatever. I'm working on that now.

• embeddings and vector database support. My hope is that the Daemon genserver can have some ready-to-use chains that help you embedify and store your data sources.

Chains:
-scrapeChain - already exists

• queryChain - chain intended to query data, a queryChain has an optional :query anchor, so it shows you what it found and asks if it looks right.
• actionChain - chain intended to perform actions, an actionChain requires an :effector anchor, so it shows you what it's going to do, warns you of potential dangers and gets permission to actually do it.

GenServers:

• Scraper - already exists
• Inquisitor - GenServer that stores and executes queryChains, inquisitors can be trusted because queryChains have no side effects, so anchors are optional.
• Daemon - GenServer that stores and executes actionChains. Daemons can't be trusted because their chains have side effects, so anchors are mandatory and have to tell you what they're going to do, warn about consequences and then get your permission before proceeding.

I already have scrapeChain and Scraper that basically work. The goal is that Daemon, Inquisitor and Scraper are high-level general-purpose AI agents that come with a toolbelt for common generic tasks so casual users can just jump right in. But people can also dig in and make their own custom chains as well.

[sethcalebweeks]

What's the best way to contribute to this project? I was about to start working on the same thing, but came across this project. I'm looking to build a conversational bot with history and the ability to use tools. We built a couple chains in Elixir using EEX to store the prompts. Would you be open to meeting with me to see if we can collaborate on this project? Thanks!

[orthagonal]

I'm definitely up for meeting contributors, I should be available any time today. What's best way for you?
I'm on both the elixir discord server and slack channels if that helps, just let me know.

[sethcalebweeks]

I couldn't figure out how to get into Slack, but I just joined Discord. My username is sethcalebweeks there. Tried finding you under the name orthogonal, but couldn't find you.

[orthagonal]

I'm HarmlessLurker#9226 on Discord, feel free to ping me!

…

On Fri, May 5, 2023 at 1:57 PM Caleb Weeks ***@***.***> wrote: I couldn't figure out how to get into Slack, but I just joined Discord. My username is sethcalebweeks there. Tried finding you under the name orthogonal, but couldn't find you. — Reply to this email directly, view it on GitHub <#1 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABRNA3EVNS44BMXEHVVXGTDXEVER5ANCNFSM6AAAAAAXWPZ5OY> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[BigSpaces]

@orthagonal and @sethcalebweeks I would humbly like to join this conversation and see if I can do something to contribute. I am quite a newbie to Elixir (and programming in general). My goal was to implement the functionalities of AI into Elixir code and this is how I came to this project (didn't want to start learning python to do a couple things, really).

It would be amazing to achieve some type of AutoGPT functionality, but with the power of Elixir.

I am in discord under @BigSpaces

[BigSpaces]

Found @orthagonal and sent a friend request :)

[cpursley]

Not sure if this is the right place to ask, but how can we use Bumblebee with scraper? I went looking through the docs (and test) and tried running a few experiments without luck. Thanks!

[orthagonal]

You should be able to back the scraper with any of the language model providers in the API wrapper, in the scraper example I think you can just replace:

openai_provider = %LangChain.Providers.OpenAI.LanguageModel{
model_name: "gpt-3.5-turbo",
max_tokens: 25,
temperature: 0.5,
n: 1
}

with (eg):

bumblebee_provider = %LangChain.Providers.Bumblebee.LanguageModel{
model_name: "distilgpt2",
max_new_tokens: 25,
temperature: 0.5
}

Bumblebee will download the Huggingface model specified by model_name to your local cache and try to run it on your GPU, so you'll need Nx and a beefy machine to run most of those. Let me know if you run into any problems.

[cpursley]

Thanks. That returned an error, I'll open an issue.