Adaptive Retrieval for Enterprise RAG
A novel approach to dynamic chunk sizing and retrieval strategies that improve accuracy by 40% on enterprise document corpora.
Enterprise RAG systems typically use fixed chunk sizes, 512 tokens, 256 tokens, regardless of document structure. But legal contracts have different density than FAQs; technical manuals have different structure than meeting notes. Fixed chunking leaves accuracy on the table.
We propose adaptive retrieval: a small classifier predicts optimal chunk size and overlap per document type, and we dynamically adjust the indexing pipeline. For retrieval, we use a query-adaptive strategy, short factual queries get smaller chunks; complex analytical queries get larger context windows.
We evaluated on 6 enterprise corpora (legal, medical, technical docs, support KBs) with human-annotated QA sets. Adaptive retrieval improved accuracy by 40% on average vs. fixed 512-token chunking, with the largest gains on document types with heterogeneous structure.
We open-source the chunking and retrieval adapters, along with benchmarks. The approach adds minimal latency (a single forward pass for the classifier) and works with existing embedding and LLM backends.
Key Findings
- 1Document-type-aware chunk sizing improves retrieval recall by 35–50% on heterogeneous corpora.
- 2Query-adaptive retrieval (short vs. long context) reduces hallucination by 25% on analytical questions.
- 3The classifier adds <20ms latency to the indexing pipeline.
Related publications.
The Architecture of Reason
An exploration of reasoning capabilities in modern language models and architectural patterns that enable multi-step logical inference.
Quantum-Classical Hybrid AI
Investigating the intersection of quantum computing and classical neural networks for combinatorial optimization problems.
Interested in collaborating?
We're always looking for research partners and academic collaborators who want to push AI forward.