01 / live demo / 116 FDA documents / 1,752 passages
It answers from the documents, or it doesn’t answer.
Ask it about FDA approvals for wound dressings. It finds the five most relevant passages in 116 official documents, answers only from those, and shows which passage each fact came from. If the documents don’t answer your question, it says so. On a regulatory question, a confident wrong answer is worse than no answer.
97 FDA clearances and 19 guidance documents · answers written by DeepSeek V3
02 / what happens to a question
Five steps, and two of them are allowed to say no.
01 / Search
Two searches through the documents: one by meaning, one by exact words. The best five passages from both are kept.
02 / Check they help
A search always finds something, useful or not. So it's asked one yes-or-no question: do these passages actually answer this? If not, it stops here.
03 / Write
The answer is written from those five passages only, and every fact has to point at the passage it came from.
04 / Check the sources
A simple check with no AI in it: every source the answer points at has to be one it was actually given. This is what stops a made-up source getting through.
05 / Retry or refuse
A failed answer gets one more try, with what was wrong. If that fails too, it refuses. No best-of-two-bad-answers.
03 / is this actually the project
The same search as the original, checked.
The real project is written in Python and is too heavy to run on a website, so this page runs a copy of its search, rewritten for the web. I tested the copy against the original on 24 questions before letting it run here. It matched on all of them.
The measurements, for anyone who wants them
24 / 24
questions tokenised identically
plus 300 chunks of corpus text
0.99999997
lowest cosine to the Python vectors
1.0 would be bit-identical
24 / 24
identical top 30, both searches
meaning and keyword, in order
24 / 24
identical top 5, every mode
dense, keyword, and fused
The search by meaning uses bge-small embeddings and the search by exact words uses BM25; thirty candidates from each are fused by rank into five. The embedding model ships at half precision, 67 MB instead of 133. That was measured before it was chosen: half precision kept every ranking identical on all 24 questions, while 8-bit, at half the size again, changed the top 30 on 23 of them.
04 / what’s next
Where I’m taking it.
Ask it harder questions.
On its own 49-question exam, each question reviewed by hand, it answered 34 of the 40 it should answer and refused all 9 it should refuse. Every miss was a refusal or half an answer, never a wrong one. Only one question needs two documents at once, so those are the next ones to write.
Search the FDA's own guidance first.
Ask what testing the FDA expects and it mostly finds manufacturers describing their own tests, so it declines, which is the right call with those passages. Putting guidance first fixes that.
Read the scanned documents.
42 of the 140 clearances I downloaded were scanned paper with no readable text. I left them out so it never answers from a blank page. Text recognition would bring them in.
It sticks to wound dressings.
Ask about anything else and it refuses, which is the point. It’s a demonstration, not regulatory advice.