You have the answer somewhere. It is in a contract signed two years ago, a supplier manual, a policy document or a long email thread. Finding it takes twenty minutes of searching and reading, and you are never sure you found the right version. Asking a question in plain words and getting the answer back, with the page it came from, is one of the most useful things AI can do for a business. It is also easy to get wrong.
Below, you will see how answers from your own documents work, how to prepare for them, and how to tell a good answer from a bad one, so you can chat with your documents and get answers you can check.
How it works, in plain words
When you add a file to a library, a system reads it and breaks it into passages. When you ask a question, it finds the passages most likely to contain the answer, hands them to an AI model along with your question, and the model writes a reply based on those passages. Showing which passages were used is what lets you check.
There are two ways the search can find passages, and a good system uses both:
- Exact words. Finds passages that contain the words you typed. Best for names, numbers, clause numbers and codes. "Clause 14.2" or "invoice 1042" should find exactly that.
- By meaning. Finds passages that talk about the same thing even in different words. Best for questions such as "can we cancel early?" when the contract says "termination before the end of the minimum term".
In nerix, the library offers both: you can search by exact words or by meaning, and the chat can answer from your files and show its sources.
Why sources matter more than fluent answers
An AI model writes fluent text, and fluent is not the same as true. A model can produce a confident, well-formed answer that is wrong, or that blends what your document says with what is generally true elsewhere. The cure is not to trust the model less in the abstract. It is to make every answer checkable.
A good answer shows where each claim came from, so you can:
- Click the source and read the original passage.
- Confirm it says what the answer says.
- See which document, which page and which version.
In nerix, each answer shows its sources, and you can open the passage and its page. If a sentence has no source, treat it as the model's own words, and check it before you rely on it.
Prepare your documents
The quality of the answer depends on the quality of the library. A small amount of preparation helps a lot.
Keep one current version
If the library holds three versions of a contract, the system may quote the old one. Remove superseded files, or name them clearly. "Supplier agreement, signed 2026" is better than "final_v3_new".
Use readable files
Text-based PDFs and Word documents work best. Scans are read too, but photos of pages at an angle or in poor light produce errors, and numbers are where errors hurt. Check important scanned documents by eye.
Organise in folders
Group files by subject: contracts, policies, supplier manuals, HR. Folders help you control what a question searches, and help people find things themselves.
Mind the limits
In nerix, each file can be up to 128 MB. Each file passes through states, Waiting, Reading, Indexing and Ready, and you can ask questions about it once it is Ready. A file that failed to read should be fixed or replaced, not ignored.
Do not upload what you should not
Be thoughtful about what goes in. Documents full of personal data should only be uploaded where you have the right to process them, and you should know where they are stored. nerix stores data in the EU, on servers in Germany, and has settings to hide names, emails and amounts before text goes to a model. Read the Security page to understand what applies to you.
Ask good questions
You get better answers from better questions. A few habits:
- Be specific. "What is the notice period in the Acme agreement?" works better than "tell me about Acme".
- Name the document or the period when it matters. "In the 2026 supplier agreement" avoids the 2024 one.
- Ask for the source. "Quote the clause" makes the model show its evidence.
- Ask one thing at a time for important matters. Combine questions only for exploration.
- Say what you will do with it. "I need to decide whether to cancel before June" helps the model pick what is relevant.
For summaries, ask for structure: "list the obligations on each side" or "list every date and deadline". Lists are easier to verify than prose.