Severn: Agent AI pipelines in Rust, made easy.

This crate aims to make writing AI agent pipelines as easy as possible, while still using idiomatic Rust syntax.

Features

• Sequential pipelines for running AI agents and feeding the results into each other
• Exposes traits you can use to make your own traits and data sources

How to use

Agents

First, you need to define a struct that will serve as your agent:

use severn::agent::Agent;

struct ExampleAgent;

impl ExampleAgent {
    pub fn new() -> Self {
        Self
    }
}

impl Agent for ExampleAgent {
    fn name(&self) -> String {
        "Example agent".into()
    }

    fn system_message(&self) -> String {
        "You are an example agent, Neo. Your job is to serve as an example for all the other agents.".to_string()
    }
}

Next, you can now use your agent by turning it into an Arc<T> and adding your agent to the pipeline:

use std::sync::Arc;
use severn::pipeline::Pipeline;
use anyhow::Result;

#[tokio::main]
async fn main() -> Result<()> {
    let example_agent = Arc::new(ExampleAgent::new());

    assert_eq!(example_agent.name(), String::from("Example agent"));

    let pipeline = Pipeline::new().add_agent(example_agent);
    
    let prompt = "What is your job, Neo?".to_string();
    
    let result = pipeline.run_pipeline(prompt).await?;
    
    println!("{result}");
}

You can also use macros to define the implementation for your agent:

#[severn(name = "Example agent",
    system_message = "You are an example for how to write macros with agents"
)]
struct ExampleAgent;

Data Sources

Severn also additionally exposes a severn::data_sources::qdrant::Qdrant struct for all of your RAG needs. Qdrant exposes a method for embedding and upserting single files into your Qdrant database. To use it, you need the qdrant feature enabled. This will also expose the qdrant_client crate as severn::qdrant_client.

Severn also additionally exposes a (WIP) HttpClient struct which allows you to add your own reqwest::Client. The reqwest crate is exposed as severn::reqwest - so you don't need to add the crate manually!

Need a custom data source? You can do exactly that! You only need to implement the DataSource trait - then you can add it to whatever pipeline you want.

Text transformation

A File trait is exposed which the Qdrant::embed_and_upsert method takes. You can either use the MarkdownFile (or CSVFile) structs, or you can additionally create your own.

Models

Currently only OpenAI is supported, but in the future support will be added for more.

Contributions

Issues and PRs are welcome. However, unless the fix is very minor (for example a documentation typo), please make sure you open an issue first! This will avoid unnecessary work if it is either not in line with the overall vision of the crate(s) or warrants more attention than a single PR.

FAQ

Why not llm-chain?

I found that the llm-chain crate was quite macro heavy and not to my taste. I'd be happy to use it if it was more idiomatic, but it isn't, so here we are.