RAG chatbot development that cites your documents

What RAG means in plain words

RAG stands for retrieval-augmented generation. When a user asks a question, the system first searches your documents for the most relevant passages. It then hands those passages to a language model with the instruction to answer from them only. The answer arrives with the passages it used.

This is different from asking a model to remember your handbook. Models do not reliably store your content, and they will fill gaps with plausible guesses. Retrieval gives the model something to read, and citations give the user something to check. For a comparison with the other route, see RAG vs fine-tuning.

Is your content ready? A five-question check

The most common reason a RAG bot disappoints is the documents, not the model. Before we quote, we ask:

• Is the information written down in text, or does it live in people's heads and chat threads?
• Is there one current version of each document, or three conflicting copies?
• Are tables, scanned PDFs and screenshots important? Those need extra processing.
• Does each document have an owner who will update it?
• Are there documents some users must not see, such as HR files or client contracts?

If the answer to the first or second question is no, the first piece of work is cleaning the content, and we will say so. A bot built on contradictory sources gives contradictory answers, and that is correct behavior.

The pipeline and the decisions inside it

Permissions deserve their own line. If a sales rep and a finance clerk should see different documents, filtering has to happen at retrieval time, before the model sees anything. Filtering the answer afterward is too late.

How we test it before anyone uses it

1. Write the question listYou and we collect 40 to 60 real questions from support inboxes, sales calls or staff chats, each with the document that holds the answer.
2. Measure retrieval aloneFor each question, did the right passage appear in the top results? If not, we fix splitting and search before touching the model.
3. Measure answersA reviewer marks each answer correct, partly correct, wrong or properly refused. We keep this score as the baseline.
4. Test the trapsQuestions with no answer in the documents, questions that contradict each other, and attempts to make the bot ignore its rules.
5. PilotA small group uses it for two weeks. Every thumbs-down becomes a new test case.

A concrete illustration

Take a manufacturer whose website lists dozens of products across styles and sizes, similar in shape to the 67-product catalogue we organized for Asian Packaging. We did not build a chatbot for that project, so treat this as an illustration only.

A buyer asks which box style suits a heavy item. The retrieval step pulls the product pages and the testing notes. The answer names two styles, links both pages and says that load rating needs a sales conversation, because the documents do not state one. That last sentence is the refusal rule working. A bot that invented a rating would be worse than no bot.

What you get, how long it takes, how we report

• The chatbot on your site, inside your help desk or in Slack or Teams
• The ingestion job that re-reads your documents on a schedule
• The question list, scores and the changes that moved them
• An admin view of unanswered questions so you can see what is missing from your content

A pilot on one document set typically takes four to six weeks, a planning estimate we confirm in the quote. Reporting is a monthly note with retrieval hit rate on the test list, the share of questions refused, the top unanswered topics and the changes we made. You get a fixed quote after a short call at contact.

When a chatbot is the wrong answer

If you have fewer than about fifty pages of stable content, a well-organized help page and good site search often beat a bot. If answers depend on live data, such as order status, you need an integration with your systems, which we describe under AI chatbot development. RAG is the right tool when the answer lives in a large, changing body of text.