Adaptive Retrieval
Definition
Deciding per query whether to retrieve, and how much — no retrieval, one retrieval step, or several rounds of retrieval and reasoning — instead of running the same RAG pipeline for everything. It is Query Routing where the routes are retrieval strategies of different cost rather than different data sources.
Why It Matters
Traffic is mixed. Simple questions (“Paris is the capital of what?”) are answered fine by the LLM alone; multi-hop ones need iterative retrieval. A fixed pipeline either overspends on the first or fails the second (Adaptive-RAG). Retrieval is not free of harm either: in RAGRoute’s MMLU experiment, retrieving from randomly chosen sources scored below not retrieving at all (71.74% vs 73.35%), because irrelevant context distracts the model.
Approaches
| Approach | Decision signal | Example |
|---|---|---|
| Entity frequency | Retrieve only when the question’s entities are rare | Mallen et al. (2023), as described in Adaptive-RAG |
| Complexity classifier | A small model predicts no / single / multi-step | Adaptive-RAG — T5-Large, labels from which strategy answered correctly |
| Query-type classifier | Factual / reasoning / summarization mapped to strategies of increasing cost | RAGRouter-Bench baselines — TF-IDF + SVM, simulated 28.1% token saving |
| ”No retrieval” as a ranked option | Rank retrievers and abstention by expected gain in answer quality | LTRR |
| Generation-time triggers | Retrieve new documents when generated tokens have low confidence | Jiang et al. (2023), as described in Adaptive-RAG; Self-RAG (Asai et al., 2024) is the other adaptive baseline there. Agents make the decision step by step — see Agentic Search |
Practical Notes
- Imperfect routers still pay. Adaptive-RAG’s classifier is only ~55% accurate, yet matches always-multi-step F1 at under half the time, because its errors mostly lean toward retrieving more than necessary.
- Cheap features go far. On RAGRouter-Bench, TF-IDF beat sentence embeddings for predicting query type.
- Query text is not the whole story. Which strategy wins also depends on corpus structure; query-only routers are a floor.
Related Concepts
- Query Routing — the general pattern
- RAG · Agentic Search — agents make the same decision step by step rather than once
- Query Classification · Query Specificity
- GraphRAG — one of the expensive strategies a router may choose