A developer benchmarked five enterprise-grade AI knowledge base frameworks, and the results will disappoint anyone chasing "more complex, more sophisticated" solutions—HyperGraphRAG, published at NeurIPS 2025, took 7 hours and 23 minutes to build its knowledge base, and scored just 0.171 on multi-hop reasoning (questions requiring multiple information fragments to be combined), a full 0.007 lower than LightRAG (a hybrid graph + vector retrieval framework), which built in 8 minutes.
What this is
An AI knowledge base feeds enterprise documents (contracts, manuals, reports) into large models so employees can retrieve information through Q&A. This year academia has bet on more complex graph structures (GraphRAG, HyperGraphRAG), claiming they capture complex relationships involving multiple entities—a hypergraph is an extended structure where a single edge can connect multiple entities simultaneously, theoretically more expressive than traditional binary relationship graphs.
The developer's test used 124 technical document chunks. The result: the theoretical advantage of hypergraphs barely materialized on real enterprise data, while build time stretched 14x longer. Worth remembering—more advanced doesn't mean more useful. This time, GraphRAG, which took 31 minutes to build, actually scored the highest multi-hop score at 0.211.
Industry view
Supporters will argue: the test used technical documents with low density of multi-entity relationships, insufficient to demonstrate hypergraph advantages; the slow build is mainly due to GLM-4's low concurrency limit—switching to GPT-4o could compress it to 30–60 minutes.
But this counterargument matters more to us: when enterprises deploy AI knowledge bases, building isn't a one-time task—every document update requires rerunning the entity extraction pipeline. 443 minutes versus 31 minutes means 14x the compute bill and human wait time. More critically, complex solutions hit token (the model's per-call word limit) truncation on dense markdown tables and lose information.
Practical conclusion for IT leaders: don't chase NeurIPS buzzwords when selecting, chase cost-effectiveness. LightRAG (8-minute build, 19-second P90 latency) is the sweet spot in this benchmark, provided you can accept its slightly elevated 10.5% refusal rate.
Impact on regular people
For enterprise IT: before deploying an AI knowledge base, calculate two costs—compute cost of initial build, and cost of every document update. Complex solutions have orders-of-magnitude higher ops and resource demands; traditional industries don't need to chase the latest.
For working professionals: when your company rolls out an AI knowledge base, getting "I don't know" answers doesn't necessarily mean the AI is bad—it might be using too complex an approach underneath and hitting word limits. Knowing this helps you communicate effectively with IT.
For the consumer market: consumer AI notes and assistant products are booming, but most use vector retrieval (the fastest-build category in this benchmark), and there's still a capability gap with enterprise-grade complex solutions—room remains in the ToB market.