RISE framework rethinks retrieval for AI agents, cutting search costs 75%
Researchers build a bounded 'interaction space' for agents, slashing latency and cost.
Current retrieval for search agents still follows non-agentic patterns: a retriever ranks documents, and the agent reads a small set. Recent direct corpus interaction (DCI) work allows agents to use shell tools like grep and file reads, but unbounded interaction doesn't scale—every broad command scans the entire corpus, causing severe latency as the corpus grows. The researchers argue that retrieval for agentic search should not just select documents for an LLM's context window, but instead construct an interaction space: a bounded subset the agent can explore with associated tools.
As a proof of concept, they introduce RISE (Retrieving Interaction SpacE). BM25 creates the boundary, and documents are pre-processed for shell-style navigation. On the BrowseComp-Plus benchmark, RISE with gpt-5.4-mini matches the pure-shell DCI baseline at 78% accuracy while costing only a quarter per query. At 1M documents, RISE-BM25 reaches 81% accuracy on gpt-5.4-mini, whereas DCI on gpt-5.4-nano degrades to 60% with 33 out of 100 wall-clock failures. This demonstrates that a well-defined interaction space dramatically improves scalability and cost-efficiency.
- RISE replaces document retrieval with a bounded interaction space using BM25, enabling efficient agent exploration.
- On BrowseComp-Plus, RISE with gpt-5.4-mini achieves 78% accuracy at one-quarter the per-query cost of pure DCI.
- At 1M documents, RISE maintains 81% accuracy while DCI drops to 60% with 33% timeout failures.
Why It Matters
Makes agentic search over large datasets practical by eliminating linear scans, cutting costs while maintaining accuracy.