A Signal65 benchmark, produced with Kamiwaza, evaluating how 91 LLMs retrieve enterprise knowledge and resist hallucination - revealing why context window size is not a reliable proxy for retrieval accuracy.
How well do today's large language models actually retrieve enterprise knowledge - and when do they hallucinate? This Signal65 report, produced in partnership with Kamiwaza, introduces RIKER (Retrieval Intelligence and Knowledge Extraction Rating), a benchmark that evaluates LLM retrieval and hallucination across realistic enterprise document sets. Testing 91 models, the research found that advertised context window size is a poor proxy for reliable retrieval: 27 models cleared 95% accuracy at 32K context, but only three held that line at 200K. Multi-document aggregation degraded more than twice as fast as single-document retrieval, while reasoning ("thinking") modes measurably improved results. Using an inverted-generation method and deterministic grading, RIKER resists the contamination and judge bias that undermine public leaderboards. Essential reading for anyone evaluating models for RAG, knowledge management, or agentic workflows. Download the full report for the complete methodology and model-by-model results.