It knows what you mean to ask,
and where the answer lives.
Before you ask it. The assistant already knows your patterns: who you are, what you work on, what happened lately and what comes next. So "What did we promise them?" lands the way it would with a colleague, and the answer comes back with proof.
When does theira notice periodb endc?
1Lease_v3 · §4 2Renewal thread or an honest "I don't have that."
An investigation, not a lookup.
The assistant doesn't meet your question cold. It knows who is asking, what they work on, what they asked before and what moved around them this week, so it starts where the answer most likely lives. And when the first pass doesn't settle it, it keeps going: searching again, reading full documents page by page, consulting a map of the corpus. Always inside the asker's permissions.
Numbered per answer; inline markers and source chips open the exact passage
Zero valid citations → an honest "I don't have that", never a guess
Filter every step, there is no unfiltered path to the index
Evidence has a rank. Answers follow it.
Every document is indexed with its situation: date, author, recipients, the entities it concerns, and its standing on an evidence ladder, an executed instrument outranks a report, a report outranks notes, notes outrank drafts. When versions of the truth collide, the answer takes the strongest, freshest evidence and cites it, so you can see exactly which document carried the day.
Dates are handled deterministically: computed offsets accompany every retrieved date, and the model is forbidden from calendar arithmetic. "The latest contract" means the latest contract.
Executed instrument > report > notes > drafts, ranked at indexing
Question intent decides the window: "current", "as of March", "ever"
People, companies and projects resolved before searching, "them" has a referent
Deterministic offsets; no model math on calendars
Conversations persist per person; follow-ups carry their context
Answers follow the asker's language, Greek, English, mixed corpora are normal
Accurate AI answers, explained
How does it know what I mean without me spelling it out?
Because it already knows your patterns. It knows who you are, what you work on, what you asked before and what changed around you lately, so it reads a question the way a colleague at the next desk would, and it already knows which corner of the company record the answer lives in.
How can AI answer from company documents without hallucinating?
It is not allowed to speak without evidence. Every claim cites a passage from your own documents, and an answer it cannot back is an answer it does not give. You can stop worrying about invented facts; the mechanics are written up in the docs.
What happens when documents contradict each other?
It happens in every company, and the assistant is built for it: newer and more authoritative documents win, an executed contract outranks a draft, and the answer cites the document that carried the day so you can check it yourself.
What if the answer is not in our documents?
You get an honest "I don't have that in the knowledge base." No answer is invented; there is no answer without a source.
Does it handle dates and "the latest version" correctly?
Yes. Dates are computed, never guessed, and "latest", "current" and "as of March" resolve against the documents' own dates. If date handling is the thing your team worries about, the docs walk through exactly how it works.
Which languages does it understand?
Ask in your language and it answers in kind; Greek, English and mixed corpora are normal. Conversations persist per person, so follow-up questions carry their context.
Ask it the way you think.
The demo corpus is full of contradictions, versions and half-named things, exactly like yours. Ask the way you'd ask a colleague and see what comes back.