trivium
Star Get started
Documentation / v0.1

Run a benchmark you can explain.

trivium gives retrieval engineers a clean way to compare BM25, dense vector search, fusion, and reranking. The workload is reproducible, the numbers are inspectable, and the limitations are part of the product.

Install from source Read the methodology

Start with the path that fits your question.

01 · Setup

Getting started

Install the checked-out source, build the corpus cache, and run the first benchmark sweep.

Open guide → 02 · Concepts

Architecture

See how documents move through embedders, retrievers, fusion, reranking, and evaluation.

Explore the system → 03 · Evidence

Methodology

Understand the corpus, metrics, latency protocol, and the claims the benchmark does not make.

Read the evidence → 04 · Build

API reference

Use the small public contracts directly in your own benchmark or retrieval service.

View the API →

Four steps from checkout to result.

  1. Install trivium from the repository so the code you run is the code you can inspect.
  2. Prepare the BEIR scifact seed and the distractor-augmented corpus cache.
  3. Run one or more registered pipeline modes at the scales you want to compare.
  4. Read the CSV rows, latency distribution, reproducibility manifest, and pre-flight result together.

The shortest useful install.

This is source-first software. A package release is not assumed; the canonical path is the repository checkout.

git clone https://github.com/sachncs/trivium.git
cd trivium
python -m pip install -e ".[dev]"
trivium-prepare --max-scales 100000
trivium-benchmark --modes bm25,vector,hybrid_rrf
What you get: a benchmark CSV with metrics, p50/p95 latency, configuration, runtime details, and the commit that produced each row.