Answers from your documents, with the clause cited
The agent searches your contracts, policies and internal guides. It answers briefly and shows the source right away: file, section, clause. If the documents don’t cover it, it says so.
Contracts, policies, guides, the wiki. Together with the document owners we drop duplicates and outdated versions, so the agent only answers from what you’ve approved.
Documents are split by section and clause, keeping clause number, date and version. That’s why the link points to clause 4.2, not to “the whole contract”.
For each question the agent first retrieves matching fragments and writes the answer from them alone. This approach is called RAG.
Every answer comes with a quote and a link to the file. Checking it takes ten seconds instead of blind trust.
If no fragment fits, the agent replies that the documents don’t cover it and offers to route the question to the person in charge.
New versions enter the base by the rule you choose: from a folder, from the wiki or after sign-off. Old versions get flagged.
One task, three different projects. Pick the one that looks like you and press “I want this” — your request arrives tagged, and we send an estimate in 1–2 days.
One team and one folder of policies that people keep asking the same questions about.
What we build
Several departments, hundreds of contracts and policies, and different people need to see different things.
What we build
Contracts contain trade secrets and personal data, and security won’t sign off on any cloud.
What we build
Pilot from RUB 300,000. Beyond that we price by task and hours.
Language models can be wrong, so we don’t ask the model to remember anything. It answers only from retrieved fragments and shows the quote. If nothing fits, the answer is “the documents don’t cover this”. Before launch we run a set of real questions from your staff and go through every miss.
Permissions are set before launch. The agent only searches documents the employee’s role allows. A sales manager won’t get a quote from an HR policy, even when asking about it directly.
We sort that out at the start. The current version is marked, old ones are removed from search or kept with a “superseded” flag. The agent states which version its answer is based on.
Yes. If contracts can’t leave the building, we put the model and the base on your servers. The architecture is designed for 152-FZ, and we work under NDA. GPU hardware is quoted separately, not hidden in a subscription.
Yes, if the scan is legible. We run a sample through text recognition first and check the quality. Unreadable pages show up right away; rescanning them beats making the agent guess.
Name the process that eats time. You get a pilot range, not a 40-slide deck.