Ayanami0730

Ayanami0730 / arag

Public

A-RAG: Agentic Retrieval-Augmented Generation via Hierarchical Retrieval Interfaces. State-of-the-art RAG framework with keyword, semantic, and chunk read tools for multi-hop QA.

agentresearchlab.orgagentsa-rag agent agentic-ai agenticrag deepresearch evaluation
172
21
100% credibility
Found Feb 06, 2026 at 59 stars 3x -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

A-RAG is an open-source framework for building AI agents that use layered search tools to retrieve and reason over document chunks for accurate question answering.

How It Works

1
🔍 Discover A-RAG

You hear about A-RAG, a smart helper that lets AI dig deep into documents to give spot-on answers to tricky questions.

2
📥 Grab sample documents

Download ready-made question sets and document pieces to try it out right away.

3
🛠️ Prepare the search magic

Quickly organize your documents so the AI can search them super effectively – it takes just moments.

4
🔗 Connect a thinking brain

Link it to an AI service that does the heavy reasoning, like a super-smart friend.

5
❓ Ask tough questions

Feed in your questions and watch the AI search, read, and piece together perfect answers.

6
📊 Check the scores

See how accurate the answers are with built-in checks that grade them automatically.

🎉 Master complex queries

Now you have a reliable sidekick for getting precise insights from any big pile of info!

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 59 to 172 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is arag?

arag is a Python framework for agentic RAG, letting LLMs autonomously handle multi-hop QA over document chunks via keyword search, semantic search, and chunk reading tools. It solves the limits of traditional RAG by giving models hierarchical retrieval control in a ReAct loop, scaling performance with stronger backbones like GPT-4o-mini or GPT-5-mini. Users get CLI scripts to build indexes, run batches on benchmarks like MuSiQue or HotpotQA, evaluate accuracy, and a simple API to query custom corpora.

Why is it gaining traction?

It crushes baselines on multi-hop benchmarks—up to 94% LLM accuracy on HotpotQA with GPT-5-mini—while using fewer tokens than GraphRAG or LinearRAG. Devs dig the quickstart: clone, uv sync, build index with Qwen embeddings, set OpenAI env vars, and batch-run. As a true agentic RAG (autonomous strategy, iterative execution), it shines for "how to build agentic RAG" experiments, adapting retrieval without rigid workflows.

Who should use this?

RAG engineers tuning multi-hop QA pipelines over enterprise docs or wikis. AI researchers benchmarking "what is agentic RAG" vs. vanilla or graph methods. Teams prototyping chunk-based generation on custom datasets, especially if exploring test-time compute scaling.

Verdict

Worth forking for agentic RAG prototypes—detailed README and evals make it dead simple to benchmark your data, despite 96 stars and 1.0% credibility signaling early alpha status. Skip for production until multi-provider support lands; otherwise, it's a solid "build a RAG GitHub" starter.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.