About
Comparisons of self-hosted knowledge base platforms are almost entirely written from vendor documentation. This one is written from deployments.
What we actually do
Install each platform on a clean host from its published image, point every one at the same embedding model, feed them the same public documents, ask the same questions, and publish everything the run produced — including what failed.
Why bother
Reading documentation tells you what a platform intends to do. Running it tells you what it does. The gap between the two turned out to be the most interesting part of this project: official compose files that do not start, installs that report success while the product shows nothing, endpoints that accept a partial update and quietly discard it. None of that appears in a feature matrix.
Some of what we found is unflattering to platforms we expected to do well, and it is published anyway. A comparison that only ever confirms what its author expected is not worth the electricity it took to run.
How to disagree with us
The harness is open source and rebuilds the corpus from the same public endpoints, so a run on your hardware is comparable to ours. If your numbers differ, that is useful information and we would like the issue. If you think a dimension is weighted wrongly, the raw per-item results are published precisely so you can recompute the table with weights you prefer.
If something here is simply wrong, tell us and we will correct it in place with the revision date updated. We would rather be corrected than be quoted incorrectly.
What this is not
- Not a review of answer quality. The language model is held constant on purpose, so what is measured is retrieval, not generation.
- Not a substitute for your own trial. Your documents are not our documents, and document type turned out to matter a great deal.
- Not a running scoreboard. Each run is a snapshot against specific versions, stated on every page.
Who operates this site and what interests are involved is set out on the disclosure page.