Research
Notes from building software that shows its work.
Benchmarks, deep dives and explainers on retrieval, citation and reading complex documents. Each states its method and its sample, and the good ones say where we lost. How the findings shape the products is on our approach page.
Start here
Start here
Benchmark
We benchmarked our own knowledge graph against “just paste it into an LLM”. It lost half the time.
Two controlled benchmarks, 217 runs, the same model for every system and no model judging another. A single prompt matched months of graph engineering on short documents; on questions across a whole corpus, top-k retrieval scored zero at counting, and only the agents could say which page an answer came from.
Read the benchmarkDeep dive
Multi-vector embeddings are great. Until they’re not.Why reading a page as many patch vectors finds the region that answers a question, why storing them naively breaks the index, and how fixed-dimensional encodings make the trade-off workable, measured on our own runs over 50 long documents.Explainer
If you don’t tell me what it is, I’ll figure it out myself: a toy JEPAA joint-embedding predictive architecture built from scratch on a toy dataset, with the statistics that keep it honest (Cramér–Wold projections, the Epps–Pulley test) explained one step at a time.