Your firm's documents

A knowledge base is a collection of documents the assistant searches. In a law firm this is where your own material goes — the contracts, opinions, memos and templates that make the difference between a generic answer and an answer about your practice. The Romanian codes and Constitutional Court decisions are already indexed and sit alongside your collections; see Legislation and case-law research.

Everything below is under Knowledge in the sidebar. If that group is not there, your role does not include document access.

What to upload

Anything the firm would otherwise search by opening files one at a time:

• Signed contracts and the templates behind them.
• Legal opinions and memos you have written.
• Internal notes, procedures and precedent files.
• Client documents you need to interrogate — statements, annexes, correspondence.

PDF, Word, Excel, PowerPoint, image and plain-text files are accepted. Duplicates are detected and skipped, so re-uploading the same file does not create a second copy.

A practical way to start: pick one matter type you handle often and upload the last few years of it. Ten well-chosen contracts answer more questions than a hundred unsorted files.

Professional secrecy still applies

Uploading puts client material into your firm’s collection on the platform. It stays within your organization and is not visible to other firms or used to train AI models — but the decision about which client material goes in, and under which engagement terms, is a professional judgement, not a technical one.

Organising collections

Knowledge → Databases lists your firm’s collections with their document and passage counts, and lets you create a new knowledge base with a Name and an optional Description.

How many you need depends on how you want to aim a question. One collection per practice area or per major client lets you point the assistant at the right material; a single collection is simpler and works well for a small firm.

The embedding settings of a collection are fixed once the first document has been processed — the panel says so — so decide on the split before you bulk-upload.

A collection’s page shows its documents with their state, and offers Upload, Edit Knowledge Base and, in a clearly marked danger zone, deletion.

Uploading

Knowledge → Documents → Upload:

1. Choose the Knowledge Base.
2. Drag files onto the drop zone, or browse.
3. Upload & Process.

Processing runs in the background: text is extracted, split into passages and indexed for both meaning-based and exact-word search. The documents list shows Pending, Ready or Failed per file, along with pages, words and the number of indexed passages — the count is how you know a file is genuinely searchable rather than merely uploaded.

A document’s own page shows its extracted text, its format and size, who uploaded it and when, View Original to open the file, and Reprocess if something went wrong. For indexed legislation it also shows detected cross-references — what the provision refers to and what references it — and links out to the official portal.

What the assistant can then do

With your documents indexed, Ask Lex can:

• Answer questions about a specific contract or opinion, citing the passage.
• Compare the same clause across several documents.
• Cross-check facts between two documents and report every mismatch.
• Pull the deadlines and time limits out of a document into a chronological list.
• Draft a memo whose analysis is anchored in the material you have.

The commands that drive those tasks are documented in Klu for legal work.

Two ways to aim a question at specific material:

• @ in the composer pins one or more documents to the question. The assistant then works from those files rather than searching the whole collection. Type after the @ to search your documents by name. Clause comparison and cross-checking rely on this.
• Dropping a file into the conversation handles a one-off: the text is used for that question without being added to a collection.

Personal data in attachments

The shield in the composer controls whether personal data in a file you attach is replaced with placeholders before the text reaches the AI model — names, addresses and other identifiers. It is on by default, and you can hover it to see its state and switch it. Legal reference material such as the indexed codes is not affected.

When personal data was replaced, the answer you read still shows the real names — the substitution protects what is sent to the model, not what you are entitled to see.

Verifying an answer

Citations are the point of the exercise, so use them:

• Numbered markers in the answer correspond to cards under Show sources (or Sources consulted).
• Each card names the document or article, the page range and the section heading where available, and shows a short excerpt.
• The link on the card opens the source — the uploaded file for your own documents, the official legislative portal for an article.

Under the answer, the grounding badge tells you what kind of support it has, and Lex warns you explicitly when the answer names a legal provision it could not substantiate. That behaviour is described in full in Legislation and case-law research.

If the assistant cannot find a basis for an answer in the selected documents, it says so rather than inventing one. An answer with no sources is a signal to check whether the relevant document is actually in a collection and finished processing.

Scope and access

Collections belong to your organization. If you work in more than one, you see the collections of the one you are currently in. The shared Romanian legislation and Constitutional Court corpora appear in the picker alongside your own — searchable by everyone in the firm, and not something anyone can alter.

Related

• Getting started walks through a first upload and a first question.
• Klu for legal work is the command reference.
• Legislation and case-law research covers the reference corpus.