Crate quorumrag 

quorumrag

Quorum-based retrieval-augmented generation.

Instead of trusting a single retriever, QuorumRAG runs several retrievers (dense embeddings at different chunk sizes, plus BM25) over the same corpus, fuses their results with reciprocal rank fusion, clusters semantically similar evidence, and then keeps only the clusters that multiple retrievers agree on (the “quorum”). The surviving evidence is ranked and turned into a context string that can be handed to a generator.

The crate talks to a local Ollama server for both embeddings and generation.

Quick start

use quorumrag::{Config, QuorumRag};

let config_str = std::fs::read_to_string("config.toml")?;
let config: Config = toml::from_str(&config_str)?;

let rag = QuorumRag::build(config).await?;
let answer = rag.answer("What is the capital of France?").await?;
println!("{answer}");

For finer control, QuorumRag::retrieve returns the ranked EvidenceClusters and the assembled context without invoking the generator.

Re-exports

pub use config::Config;

pub use config::OllamaConfig;

pub use config::RetrieverConfig;

pub use config::RetrieverType;

pub use embedding::EmbeddingClient;

pub use generation::Generator;

pub use models::Candidate;

pub use models::Chunk;

pub use models::EvidenceCluster;

pub use models::Query;

Modules

clustering

config

context

corpus

embedding

evaluation

generation

models

quorum

ranking

retrievers

Structs

QuorumRag

A fully indexed QuorumRAG pipeline: retrievers, an embedder, and a generator wired together from a Config.

RetrievalResult

The outcome of a retrieval: the assembled context plus the structured evidence behind it.

Functions

embed_chunks_parallel

Embed chunks in batches of Config::embed_batch, embedding each batch concurrently against the Ollama server at base_url.