ERAST: Efficient Retrieval-Augmented Search Tool

This repository contains code for training ERAST and running its localized version.

---

ERAST is a high-performance framework for large-scale homology searches across ~1 billion biological sequences — hosted in the largest vector database to date.
By integrating state-of-the-art large language models with vector database technology, ERAST delivers both speed and precision in detecting homologous biological sequences.

Key Features

• Multi-stage optimization — Pre-retrieval, retrieval, and post-retrieval stages to boost accuracy.
• Broad applicability — Supports both nucleotide and protein sequences, unlike most protein-only tools.
• Scalable performance — Efficiently handles billion-scale datasets without compromising search quality.

---

Repository Structure

1. Training Pipeline

Train your own ERAST models:
ERAST training

2. Localized Version

Run ERAST locally for custom datasets:
Pipeline

---

Environment Setup

Requirements:

• Python 3.8+
• HMMER (for Pfam scanning)
• Python packages:

pip install numpy torch datasets transformers

---

Access the ERAST Vector Database

The public ERAST vector database can be accessed here:
https://ai4s.tencent.com/erast