01 / 03

An answer has to come from somewhere

A colleague answering a delivery question would usually open the order record and the latest warehouse update. An assistant needs access to the relevant information too.

In Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, Patrick Lewis and colleagues combined a language model with retrieved text. Their research used a Wikipedia index and evaluated tasks including question answering. The work helps explain an approach now commonly called RAG: retrieve relevant material and use it when generating an answer.

It does not mean that giving a model documents makes every answer correct. Finding the right material and using it properly remain things to test.

02 / 03

The company version needs care

Our practical starting point is a small collection of current, owned documents. Each should have someone responsible for it and a clear date. Remove conflicting versions before asking an assistant to choose between them.

Keep permissions attached to the material. An assistant should only retrieve information the person asking is allowed to see. A convenient search box should not quietly become a way around access rules.

For an operational answer, make the underlying record easy to open. A source link lets someone check the date, the context and the wording. It is part of doing the work, especially when the answer will affect a customer or an order.

03 / 03

Test the gaps

Include questions with no answer in the approved material. The assistant should say what is missing and give the person a sensible next step. Add questions where an old policy and a new one disagree.

When a response is wrong, check two things separately: did the system find the right source, and did it represent that source correctly? Each failure needs a different fix.

These are implementation choices we would examine in a company setting. The original paper gives the underlying idea; the quality of a working system still depends on the records, permissions and checks around it.