A research team has built a conversational business intelligence system that handles both structured databases and unstructured documents in a single interface.
COGNI, described in a new arXiv paper, layers four components to handle mixed-modality enterprise queries. An indexing layer uses different chunking strategies depending on content type — plain text gets recursive chunking, while tables and charts get hierarchical treatment — hitting 88.3% on an internal benchmark. A fine-tuned Qwen-2.5-1.5B-Instruct model then routes each query, deciding both which modality to hit and how complex the question is, at 93.8% accuracy and roughly one-seventh the cost of a frontier model. From there, the system runs either a self-correcting NL2SQL agent (93.9% on the G-Eval SQL generation benchmark) or a recursive retrieval pipeline that handles multi-hop document synthesis at 91.0%. A caching layer validates query equivalence across multiple dimensions, claiming zero false cache hits and an 8.4x latency reduction.
The problem COGNI solves is real and underserved. Enterprise BI tools have long forced a choice: you get a SQL interface for your data warehouse or a document search for your slide decks, not one system that reasons across both. Every analytics team that has ever watched a business user paste the wrong query into the wrong tool knows the friction. A router that correctly classifies modality 93.8% of the time at a fraction of frontier-model cost is a meaningful engineering result, not just a research curiosity.
That said, the 88.3% indexing benchmark is self-reported on an internal dataset, and production BI systems live or die on the edge cases that benchmarks miss — the malformed slide, the ambiguous metric definition, the warehouse schema that changed last Tuesday.