An assistant on your documents
The same question comes up thirty times a month, and the answer already exists in a procedure nobody can find. An assistant connected to your own documents answers in its place, and cites the document its answer comes from.
A general-purpose assistant does not know your prices, your warranty periods or the name of your stainless steel screw supplier. Asked about them, it invents a plausible answer, which is worse than no answer at all. The technique that fixes this has a name, RAG, for retrieval augmented generation. The principle fits in one sentence: the model is never asked to know, it is asked to look. It finds the relevant passages in your documents, indexed in a vector database, then writes from those passages only.
What separates a demo from a working system
This is no longer a feat, it is an established building block. What separates a serious implementation from a demo comes down to three far less spectacular things: how well the documents are split up, how access rights are respected, and what the assistant does when it finds nothing.
On that last point we are inflexible. An assistant that finds nothing must say so and stop. We would rather have a system that returns “I do not have that information in the documents provided” than one that is right nine times out of ten, because nobody ever notices the tenth.
What a deployment covers
- Your sources taken on: office documents, PDFs, internal pages, knowledge base.
- Splitting and indexing, redone every time a document is updated.
- An answer citing its source, with a link to the exact passage.
- Access rights respected: an employee receives only what they may read.
- An explicit refusal when the information is not in the sources.
- A log of the questions asked, showing what your teams are really looking for.
The questions asked are often worth more than the answers
That last point is often worth more than the assistant itself. After a month, the list of questions asked says exactly which procedures are unreadable and which information is missing. The documentation then corrects itself, before anyone has even measured the time saved.
Two natural extensions. If the documents cannot leave the company, the whole thing is installed on your own servers. And if you expect the assistant to act, to open a ticket or send a reply rather than dictate one to you, what you are looking for is an agent.