A-RAG: Agentic RAG via Hierarchical Retrieval (2026)

Free

Hierarchical retrieval interfaces enabling agents to dynamically navigate multi-level knowledge structures

FreeFree tier

Type

Open Source

Visit Website GitHub (40 stars)

About A-RAG: Agentic RAG via Hierarchical Retrieval (2026)

A-RAG is an agentic retrieval-augmented generation (RAG) framework that exposes hierarchical retrieval interfaces—keyword search, semantic search, and chunk read—directly to the language model. Unlike traditional single-shot or predefined workflow RAG approaches, A-RAG allows the model to adaptively search and retrieve information across multiple granularities, effectively leveraging its reasoning and tool-use capabilities. Experiments on multiple open-domain QA benchmarks show that A-RAG consistently outperforms existing RAG methods while using comparable or fewer retrieved tokens. The framework scales with model size and test-time compute, and the authors plan to release code and an evaluation suite to facilitate further research.

Key Features

Exposes hierarchical retrieval interfaces: keyword search, semantic search, and chunk read

Enables the language model to adaptively search and retrieve across multiple granularities

Consistently outperforms existing RAG approaches on open-domain QA benchmarks

Uses comparable or fewer retrieved tokens than prior methods

Scales effectively with larger models and increased test-time compute

Planned open-source release of code and evaluation suite

Pros & Cons

Pros

Outperforms existing RAG methods on multiple benchmarks
Lower token usage compared to baseline approaches
Adaptive and dynamic retrieval strategy leverages model reasoning capabilities
Scales well with model size and test-time compute
Open-source code release enables community experimentation and improvement

Cons

Currently a research paper; not a production-ready product
Evaluation limited to open-domain QA benchmarks; generalization to other tasks not fully tested
Requires access to capable language models with strong reasoning and tool-use abilities
Dependence on hierarchical retrieval tools may add complexity to implementation

Best For

Open-domain question answeringKnowledge-intensive information retrieval tasksComplex reasoning tasks requiring adaptive retrieval strategies

FAQ

What is A-RAG?

A-RAG is an agentic retrieval-augmented generation framework that gives the language model direct access to hierarchical retrieval interfaces—keyword search, semantic search, and chunk read—so it can adaptively retrieve information at different granularities.

How does A-RAG differ from traditional RAG?

Traditional RAG typically uses single-shot passage retrieval or a predefined workflow. A-RAG lets the model participate in retrieval decisions, dynamically choosing which tool to use based on the query.

What retrieval tools does A-RAG provide?

A-RAG provides three tools: keyword search, semantic search, and chunk read, covering different levels of granularity and search strategies.

What benchmarks were used to evaluate A-RAG?

The paper evaluates A-RAG on multiple open-domain question answering benchmarks.

Is the code for A-RAG available?

The authors state they will release the code and evaluation suite to facilitate future research.