What to load
Start with the documents lawyers return to most often. This is know-how that now lives in email and on the drives of individual staff, and leaves with them.
1. Opinions and memoranda: the firm’s position on a question and its arguments.
2. Templates and standard contracts, with the history of edits and comments on disputed clauses.
3. Claims, replies and appeals from closed cases, together with the final court decisions.
4. Firm policies: conflict of interest checks, document formatting, handling deadlines.
5. Collections of case law that lawyers have already put together for their practice areas.
What not to load at the first stage: drafts without a final version and correspondence without a conclusion. They clutter answers with conflicting positions.
Almost all of these materials contain information about clients. Under Article 8 of Russia’s Federal Law on Advocacy, attorney-client privilege covers any information related to legal help for a client, and only the client can waive it. So the knowledge base lives inside the firm’s own systems, not in a public service.
An answer with a link to the document
The main rule of a knowledge base: no link, no answer. To the question “what position did the firm take on penalties in supply contracts”, the system answers in two or three paragraphs and puts the document, section and date under each point.
The link is not decoration. In May 2024, Stanford RegLab and HAI tested legal AI research tools: Lexis+ AI and Ask Practical Law AI got more than 17% of answers wrong, and Westlaw AI-Assisted Research more than 34%.
In an earlier study by the same authors, general chatbots got 58–82% of legal questions wrong. The gap is large: tools that search a document base made fewer errors than chatbots that answer from the model’s memory.
Search across your own documents lowers the risk but does not remove it: a model can retell a found passage inaccurately. So the lawyer opens the document from the link and sees the quote in its original context.
If the base has no suitable document, the system is set to answer “the firm’s documents have no answer” instead of making one up from general knowledge. Such a question goes to the list of gaps in the base.
Access rights by case
Access to the knowledge base mirrors access to cases. A lawyer sees in answers only documents from the cases and clients they are cleared for, and a partner sees all cases of their group.
This is needed both for attorney-client privilege and for conflicts of interest. If the firm has split teams by case, the AI must not carry information from one case into an answer about another.
1. Rights come from the firm’s case management system, not set up by hand a second time.
2. A closed or sensitive case can be switched to partners-only mode.
3. Every query and every opened document is written to a log.
In our rollouts, we process personal data on servers in Russia under 152-FZ, Russia’s personal data law. The text does not go abroad or to services that have no contract for processing on the firm’s behalf.
The AI has its own account and rights only for its process. The system, the data and the access stay with the firm, and the firm’s administrator can revoke any access.
Keeping the base current
A knowledge base is useful while it is fresh. A new document enters it when a lawyer saves the final version to the case folder or the document management system, with no separate upload.
Outdated documents are not deleted but marked. An opinion on an earlier version of a law stays in the archive, but in an answer the system shows its date and warns that the position may have changed.
1. Once a week a report arrives: which documents were added and which questions the base did not answer.
2. The person in charge of the base, usually a senior lawyer or the head of a practice, decides which gaps to close with a new opinion.
3. Lawyers’ marks on answers (“inaccurate”, “wrong document”) go into tuning the search.
This cycle keeps quality under control: the errors Stanford found in ready-made tools show up in your own base through lawyers’ marks and are fixed at the source, not in every answer.
How it works technically
Technically, this is search with answer generation. Documents are split into passages, a numeric representation of meaning is built for each, and the passages are stored in a vector database.
For a lawyer’s question, the system finds passages close in meaning, taking access rights into account, and passes only those to the model. The model writes an answer from these passages and adds links to the documents.
The parts of the solution: a server or cloud in Russia, a language model, a vector database of documents, a connection to the case management system for access rights, and a query log.
The firm decides where the model runs. It can run on the firm’s own server or in a cloud in Russia, and you can change the model without rebuilding the base and the access rules.
A knowledge base often becomes the first module of an AI assistant for lawyers: the same search across the firm’s documents later prepares contract drafts and case notes. The decision and the signature stay with the lawyer.