Nine stages between a question and a sale, none of them hidden
A shopper's question becomes a recommendation through nine stages. On most platforms you can see the first and the last. When a wrong product or a wrong price comes out somewhere in between, and sooner or later it will, the difference between a five-minute fix and a week of guessing is whether you can look at the stage that produced it.
This is what ChatWidget calls a RAG chatbot: retrieval-augmented generation. Instead of answering from whatever a language model remembers from its training, it first retrieves passages from your own content and then writes the answer from those passages alone.
-
Ingest
Your content arrives three ways: a crawl of your website (following your sitemap, up to 5,000 pages), uploaded files (PDF, plain text, Markdown, HTML, CSV and Excel spreadsheets) and your product catalogue. You see what was fetched, what was skipped and why: a 404, a robots rule, a page behind a login. Sources can be refreshed by hand or on a schedule: hourly, daily or weekly.
-
Chunk
Documents are split into passages small enough to retrieve one at a time, following the structure of the page rather than a fixed character count, so a heading stays attached to the paragraph it introduces. A product becomes one passage carrying its name, price, stock, image and link.
-
Embed
Each passage is turned into an embedding, a list of numbers that places text with similar meaning close together. That is what lets “something to fry snacks in” find your groundnut oil.
-
Retrieve
For every question, the hundred closest passages are gathered, adjusted for exact keyword matches and for catalogue entries, and the best eight go forward.
-
Threshold
Each candidate carries a score. Passages below the threshold are dropped and recorded as dropped. This single line decides whether the assistant answers or declines, and it is the line our trace draws.
-
Assemble
The surviving passages are put together with your verified answers, your rules and your tone, and with a plain instruction: answer only from this, and say so when it is not enough.
-
Answer
A language model writes the reply. You choose the model in your settings; for questions about price and stock, the assistant calls your store live rather than trusting the stored snapshot.
-
Check
Before the answer is final, every image is checked against your real catalogue photos and every link into your own site is checked against pages the assistant has actually read. An invented link loses its link; an invented image is replaced or removed.
-
Trace
The answer is stored with everything above: the candidates, their scores, what was used and what was dropped. It sits beside the conversation in your console, and your playground shows the same view for any test question, with a diagnosis of what went wrong.
Questions
How does a RAG chatbot answer a question?
It retrieves the passages from your content most likely to hold the answer, scores each one, keeps those above a threshold, and asks the model to answer from those and nothing else. Every stage is shown in the trace.
What happens when nothing matches?
The assistant says it does not know and offers a person. The question is saved in your review queue with what was retrieved and why it was not enough.
Do I have to write the answers myself?
No. It reads your website, your catalogue and any files you upload. Verified answers are for the questions that must always be answered one way.
See a sale, traced
A real question, the product it found, and every candidate it read to answer, with the line that decided what reached the shopper.