AgentLex.si
An intelligence you cannot check is not an intelligence you can use. AgentLex is our answer: a research agent that reads your documents and will not say anything it cannot show you on the page.
01The name
It does the work a researcher does: decides what to look for, searches, reads where a passage is cut short, and stops when the evidence is in hand. You ask; it goes and finds.
Both meanings are the point. It works with words, exactly as they were written, and it works under a rule it never breaks: nothing reaches you that is not on a page.
Our bet about where AI is going. As models grow more capable, the question stops being "how smart is it?" and becomes "can I check it?" We build for that question.
02Why we built it
AgentLex began as the engine inside KhateebNotes, a research assistant for imams, teachers and students of knowledge. They prepare sermons and lessons from a library of over ten thousand classical Arabic books, and for them a misquoted text is not a small error. A sentence attributed to the wrong scholar, a word changed in a quotation, a reference to a page that does not say it: each one is a mistake their audience may repeat for years.
General chatbots were no help. They answered fluently and confidently, and they invented sources. So we set ourselves one rule and built everything around it: every word an answer quotes must be found on the page it cites, letter for letter, and checked by code, not by the model.
Holding to that rule taught us most of what is in AgentLex. Books had to be cut the way their authors printed them, with footnotes kept apart and a narration kept whole with its chain. Whole pages turned out to be better than small chunks for finding the right place. A reranker mattered more than anything else we tried. And the answer had to be written only from what the research found, so that when the library is silent, the answer says so.
Then we noticed the rule was not special to scholarship. A clinician needs the note where the dose was changed, not a summary of it. A lawyer needs the clause as it was enacted, with its place in the statute. An archivist needs the letter and the date. Anywhere a wrong page is a wrong answer, the same standard applies. So we rebuilt the engine as a service anyone can use on their own documents, and gave it a name of its own.
Every answer, sewn to its page.
03What we mean by superintelligence
We do not claim AgentLex is superintelligent. The name is a direction, not a boast.
Language models are getting better at reading, reasoning and writing every few months, and they will soon read a whole library faster than any person could read a chapter. That is useful only if a person can still check what they say. An answer that is brilliant but unverifiable is worse than no answer in a hospital or a courtroom, because it will be believed.
So our idea of superintelligence is not a model that knows more. It is a system that can read everything, and still show its work: every claim tied to its source, every quotation proven to be on its page, every gap admitted. The more capable the model inside it becomes, the more that discipline matters, and the more an expert can rely on it to do the reading while they do the judging.
04The rules it keeps
No answer without its page.
Every claim cites a page from your documents. Nothing is supplied from the model's memory.
Quote exactly, or not at all.
Every quotation is looked up on the page it cites, letter for letter, by code. One that is not there is reported, never kept.
Judge every claim.
Each sentence of an answer is checked against the sources it cites and labeled supported, partial, unsupported or uncited, so you can see which ones stand.
Admit what is missing.
When the documents do not answer part of a question, the answer says so instead of filling the gap.
Keep people's names out of the model.
In a protected collection, names, record numbers and phone numbers are masked before any text leaves the server, and restored only when the answer comes back.
Measure, don't describe.
Every setting was chosen by a race on real questions, graded blind. You can run the same race on your own documents.
Your library, not ours.
Documents are returned only to the account that indexed them, encrypted at rest, and never used to train a model.
05How it works
Underneath are a few unglamorous decisions that make the difference. Whole pages are the unit we search, because on our own tests they found the right place far more often than small fragments (nDCG@10 of 0.81 against 0.69). Keyword search counts for half of what meaning-based search counts for, and a cross-encoder rereads the top thirty results; on our legal benchmark that step alone lifted the score from 0.73 to 0.95. The question is searched before the model says a word, so sources appear in under a second.
For records that carry names, the software follows the technical safeguards of HIPAA's Security Rule: encryption at rest, an audit trail of every access, identifiers masked before any model call, data sent only to vendors under a business associate agreement, and retention you control. The features page walks through every part with examples, and the API reference has every field.
06What it is not
It is not an oracle. It knows only what is in the documents you give it, and it says so.
It does not replace the expert. It does the reading: finding the note, the clause, the passage, and putting it in front of the clinician, the lawyer or the scholar with the page attached. The judgment stays with them.
It is not finished. We publish what we measure, including where it falls short, and we keep racing every change against the last.
Bring a library. Ask it anything.
AgentLex is built by the team behind KhateebNotes. For a dedicated deployment, a business associate agreement, or just to talk, write to us.