Tool-Schema Compression Enables Agentic RAG Under Constrained Context Budgets
The paper discusses the challenges faced by agentic RAG systems due to tool schemas consuming context windows needed for retrieval-augmented generation. It presents a systematic study evaluating various models and the impact of tool-schema compression on performance. The findings indicate that compressed schemas significantly improve functionality in constrained contexts, establishing their necessity for effective deployments.
- ▪Agentic RAG systems struggle with resource conflicts as tool schemas occupy the same context window required for generation.
- ▪The study evaluates 14 models and shows that applying tool-schema compression can restore functionality in limited context budgets.
- ▪Compressed schemas lead to a significant increase in exact-match performance, particularly in scenarios where tool definitions overflow the context window.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Software Engineering arXiv:2605.26165 (cs) [Submitted on 24 May 2026] Title:Tool-Schema Compression Enables Agentic RAG Under Constrained Context Budgets Authors:Furkan Sakizli View a PDF of the paper titled Tool-Schema Compression Enables Agentic RAG Under Constrained Context Budgets, by Furkan Sakizli View PDF HTML (experimental) Abstract:Agentic RAG systems that equip language models with dozens to hundreds of tool definitions face a critical resource conflict: tool schemas consume the same context window needed for retrieval-augmented generation.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
