Ask

Ask about the work on this site

A question goes to the 32 passages this site already publishes about its own projects. You get back a verbatim quote, a link to the page it came from, and the scores that put it first. When nothing in the corpus matches, or the query tries to rewrite the retriever's instructions, it says so and stops.

Nothing is sent anywhere: the corpus is built from this site at build time and the retrieval runs in your browser.

Retrieval evaluation

12/12 questions cite the expected page first

A fixed set of questions, each with the page a correct retriever must surface as its top citation. It runs at build time over this corpus, so a content edit that breaks retrieval fails the release instead of quietly degrading the answers.

QuestionExpected citationTop result
how does the abstention policy work?/evaluationpass
what is the top-1 accuracy on the held-out images?/evaluationpass
how many test images abstain at the production thresholds?/evaluationpass
which runtime is faster, pytorch or onnxruntime?/benchmarkspass
what tolerance does the probability parity check use?/benchmarkspass
how do I try the api with my own photo?/demopass
what happens to a blurry or badly framed image?/demopass
what are the limitations of leafsentry?/projects/leafsentry-aipass
how was the simbk audit carried out?/projects/simbkpass
what did you contribute personally to the fish encyclopedia?/projects/kenalinikanpass
how does the fish season map work?/projects/kenalinikanpass
how are student counselling records kept isolated by role?/projects/simbkpass

How it works

The corpus is assembled at build time from the case studies, the writing, and the pinned evaluation and benchmark reports, so it is the same data the other pages render. A query is split into terms, each passage is scored with a BM25-style weighting (term frequency, length normalisation, inverse document frequency), and the top three come back with their scores and the terms that matched.

The answer is the sentence from the best passage that shares the most terms with the query, checked as a substring of its source before it renders, so a citation cannot point at text the source does not contain. Nothing gets paraphrased. There is no language model in this loop and therefore no rewrite to trust; you are reading the source line itself.

Empty queries, queries over 500 characters, instruction-override attempts, and queries with no lexical overlap come back as a refusal with a reason code attached, which is the part a retrieval system usually hides.

The limits are worth stating because they are easy to forget while typing. Lexical matching misses synonyms: "how fast" will not find "latency" unless the passage says it. It cannot reason across passages, count, or compare beyond what it retrieves, and it knows only this site, so questions about the wider world are declined. All of it runs in your browser; nothing is logged or sent anywhere.

The evaluation behind the numbers it quotes lives on the evaluation page, the runtime comparison on benchmarks, and the demo runs a live inference call.