RQ-RAG: Learning to Refine Queries for Retrieval Augmented Generation
- Paper: RQ-RAG: Learning to Refine Queries for Retrieval Augmented Generation
- Source: https://arxiv.org/abs/2404.00610
- Date: March 2024 (arXiv)
- Category: RAG / Information Retrieval
- Explicit query refinement: The model learns to rewrite, decompose, and route queries rather than passing the raw text to the retriever.
- Learned retrieval decisions: Instead of always retrieving, the model is trained to judge when external retrieval is necessary and when it is not.
- System framing: The work sits at the intersection of RAG and large-scale search/recommendation, where retrieval, ranking, and generation responsibilities must be redistributed in the LLM era.
- Methodological pipeline: Query/document encoding → retrieval and reranking modules → training (supervised fine-tuning and related strategies) → inference with iterative retrieval and budget control.
- Graph RAG surveys (arXiv:2501.13958)
- Agentic RAG (arXiv:2501.09136)
- RAFT: domain-specific RAG adaptation
- RAG vs. GraphRAG systematic evaluation (arXiv:2502.11371)
- RAG — Retrieval-Augmented Generation
- nDCG — Normalized Discounted Cumulative Gain, a ranking quality metric
- Agentic Search — Modeling search as sequential decision-making with tool calls
- Gen-IR — Generative Information Retrieval
Overview
RQ-RAG addresses a core weakness of conventional retrieval-augmented generation: the model retrieves once using the raw user query, even when the query is ambiguous, multi-faceted, or already answerable from the model's internal knowledge. The paper proposes training the LLM to refine queries actively — deciding when to search, decomposing complex requests into sub-queries, and rewriting ambiguous inputs — so that retrieval better matches the user's true intent.
Key points
Context in the RAG landscape
RQ-RAG belongs to a growing line of work that makes retrieval itself a *learnable* behavior rather than a fixed pipeline stage. Related entries in this collection include:
Practical takeaways for engineers
| Concern | Guidance | |---------|----------| | Latency | Budget retrieval steps; use cascades with early stopping and query caching | | Quality | Verify offline gains translate to online satisfaction; audit citations | | Safety | Filter open-web retrieval for poisoning and bias; whitelist sources | | Cost | Route simple queries to smaller models; mix sparse and dense retrieval |
Caveats
This post is based on the paper's public metadata and abstract. Quantitative experimental results, datasets, and baselines should be verified directly against the original PDF before citing specific numbers.