Research
What we measure, and what we can't yet show.
Two write-ups. One is a reliability result we are pleased with. The other is an evaluation result that does not support the thing we sell, and we publish it anyway.
13 August 2026 · 8 min
What we measured, and what we haven't
The evaluation harness, the refusal controls, the graph participation rate of 201 in 231 runs, and the plain statement that graph retrieval has not been shown to improve accuracy over plain RAG on our question sets.
Read15 August 2026 · 9 min
From nine documents lost to zero
Our extraction pipeline lost documents silently and its worst step consumed 348% of the load-balancer timeout. What broke, what fixed it, and why the p90 improvement is smaller than it looks.
Read
The result we would rather not have.
Graph-derived retrieval participates in rank fusion on 201 of 231 question runs. On the question sets we can score, it has not answered a single question that vector-only retrieval missed — and one question went the other way.
We publish it because a category where everyone posts a favourable chart is a category where charts carry no information, and because someone would eventually run the comparison themselves.
What we can show
Reliability, measured on our own pipeline.
Before-and-after on the same system, about throughput and loss rather than answer quality.
- Documents lost
- 9 in 205improved tozero
- Extraction p90
- ~103 simproved to5.2 s
- Throughput
- 3.0 / minimproved to17.1 / min
- Worst pipeline step
- 348% of budgetimproved to4%
Across the same corpus, after moving orchestration onto durable steps with per-step retries.
Per document, sharded and run concurrently under a per-organisation ceiling.
Documents fully processed per minute, same corpus, same hardware shape.
As a share of the load balancer timeout. Above 100% the request dies before it finishes.
Run your own evaluation.
The free tier is enough to build a question set over your own content and measure it — including the refusal controls, which are the part that matters.
- no card required
- first workspace is free
- one POST to ingest