Seedframe 🌱

A clean, macro-driven Rust library for building LLM applications.

Features

• Declarative API through straight forward proc-macros
• Modular Architecture with clearly defined components:
  • Loaders: Data ingestion from various sources (files, APIs, etc.)
  • Vector Stores: Embedding storage and retrieval (In-memory, Redis, etc.)
  • Embedders: Text embedding providers
  • LLM Clients: Unified interface for different LLM providers
  • Tools: Function calling abstractions with state management and automatic documentation
  • Extractors: Structured output generation from LLM responses

Installation

Add to your Cargo.toml:

[dependencies]
seedframe = "0.1"
tokio = "1.44"
async-trait = "0.1"
# If you'll be using Extractors or custom types as tool-call arguments
schemars = "0.8.22"
serde = { version = "1.0", features = ["derive"] }
serde_json = "1.0"
dashmap = "6.1"

Usage

This library is in early stages and its API is subject to change.

Check out the examples directory for detailed usage demos.

Tool calling and structured extraction

The tool proc-macro is responsible for declaring tool calls, and the tools attribute on the client proc-macro attaches them to the client. The macro parses the descriptions for the function and it's arguments from the doc comments, its an error not to document the function or not to document every argument except the state arguments

You could also extract structured output from the llms, the target types need to implement the Extractor, schemars::JsonSchema and the serde::Deserialize traits. Like the tools the description for the type and for it's fields will get extracted from the docs and get passed to the llm, but its not an error to leave them undocumented.

useseedframe::prelude::*;useseedframe::providers::completions::OpenAI;#[client(provider = "OpenAI", tools("analyze"))]struct ToolClient;/// Perform sentiment analysis on text
/// # Arguments
/// * `text`: Input text to analyze
/// * `language`: Language of the text (ISO 639-1)
#[tool]fn analyze(text: String,language: String)-> String {todo!("implementation");}#[derive(Extractor)]struct PersonData{/// Age in years
age: u8,/// Email address
email: String
}#[tokio::main]asyncfn main()-> Result<()>{letmutclient=ToolClient::build("You're a data analyst".to_string()).await.with_state(AppState::new())?;// Tool call
client.prompt("Analyze this: 'I love Rust!' (en)").send().await?;// Structured extraction
letperson=client.prompt("John is 30, email john@example.com").extract::<PersonData>().await?;}

Building a simple RAG

useseedframe::prelude::*;useseedframe::providers::{completions::OpenAI,embeddings::OpenAIEmbedding};useseedframe::vector_store::InMemoryVectorStore;// Declare file loader that doesnt check for updates, loading files that match the glob pattern
#[loader(kind = "FileOnceLoader", path = "/tmp/data/**/*.txt")]pubstruct MyLoader;#[vector_store(store = "InMemoryVectorStore")]pubstruct MyVectorStore;#[embedder(provider = "OpenAIEmbedding")]struct MyEmbedder{#[vector_store]my_vector_store: MyVectorStore,#[loader]my_loader: MyLoader,}#[client(provider = "OpenAI", config = r#"{"model": "gpt-4o-mini"}"#)]struct MyClient{#[embedder]my_embedder: MyEmbedder,}#[tokio::main]asyncfn main(){letmutclient=MyClient::build("You are a helpful assistant".to_string()).await;tokio::time::sleep(Duration::from_secs(5)).await;letresponse=client.prompt("Explain quantum computing").send().await.unwrap();}

Built-in Components

Loaders

• FileOnceLoader - Load files once using glob patterns
• FileUpdatingLoader - Load files and watch for changes

Vector Stores

• InMemoryVectorStore - Simple in-memory vector storage implementation

Completion Providers

• OpenAI - OpenAI API integration
• Deepseek - Deepseek API integration
• Xai - Xai's API integration

Embeddings

• OpenAI - OpenAI embeddings API integration

Integrations

SeedFrame supports extending functionality through external crates. To create an integration all thats needed is to provide a type that implements the relevant trait from seedframe (Loader, CompletionModel, etc.). You can use the following crates as inspiration if you want to write an integration crate of your own.

Completion Providers

• seedframe_anthropic - Anthropic API integration

Embedding Providers

• seedframe_voyageai - VoyageAI embeddings

Vector Stores

• seedframe_pinecone - Pinecone vector database integration

Loaders

• seedframe_webscraper - Web scraping using scraper-rs

If you wrote an integration crate, please update this list and submit a PR.

Contributing

All contributions as welcome! This library could use support for more loaders, vector stores, providers ... so don't shy away from helping!

⭐ Leave a Star!

If you find seedframe helpful or interesting, please consider giving it a star so more people get to see it!