How gair-nlp/asi-arch works
Project Type and Purpose
• This repository implements a full-stack, multi-agent autonomous research framework for discovering novel linear‐attention architectures.
• It consists of three loosely coupled subsystems:
1. Autonomous Architecture Discovery Pipeline (pipeline/)
2. Architecture Database Service (database/)
3. Cognition Base Knowledge Service (cognition_base/)
• The goal is to continuously generate, evaluate, analyze, and store new model architectures guided by scientific literature.
File Structure
• Top-level:
• README.md, requirements.txt, LICENSE, images/, index.html
• cognition_base/ – RAG knowledge service + Docker compose + raw JSON corpus
• database/ – MongoDB + FastAPI service + scripts
• pipeline/ – Four agent modules (evolve, eval, analyse, plus helpers) and orchestration script
• data/, resources/ – sample data dumps and documentation
• Each subfolder includes its own requirements.txt or Docker compose for independent deployment.
README for Architecture Insights
• The README describes a continuous loop:
Sample candidate → Evolve (Planner, Checker, Deduper) → Evaluate (Trainer, Debugger) → Analyse (Analyzer) → Update database.
• Two backend services (Database & Cognition Base) run as containers or Python APIs.
• Pipeline agents communicate with both services via REST endpoints.
Detail
• Both Docker Compose orchestration and standalone Python APIs.
• Technology stack (Python, FastAPI, Flask, FAISS, MongoDB)
• Workflow steps (Sample → Evolve → Evaluate → Analyse → Update)