Every recommendation shows its work
When a shopper is shown the wrong product or told the wrong price, the question every merchant asks is “why?”. On most platforms nobody can answer it. ChatWidget stores the answer to that question with every reply: the product it found, the passages it read, what each one scored, the ones it was not allowed to use and why.
Do you have a 28cm dosa tawa, and is it in stock?
Yes — one option, in stock today:
Sourced from this store
Retrieved · 3 candidates
Hard-anodised dosa tawa 28cm · ₹949 · in stock · dispatched in 2 days.
Orders ship within 2 business days across India.
Our spring sale runs until the end of April…
dropped · below threshold 0.32
What the trace shows
For every answer, up to twenty candidate passages, each with its source page, a preview of its text and its score. A line marks the threshold. Candidates above it were used; candidates below it are drawn dashed and marked as dropped. The trace sits beside the conversation in your console, and the playground shows the same view for any question you type.
Three causes of a chatbot hallucination, three different fixes
When a shopper is shown the wrong product, the report you receive is “chatbot hallucination”. Most of the time it is not. The trace tells three very different failures apart.
-
The right passage was never retrieved
It is not in the candidate list at all. The fix is in your content: the page was never added, was blocked from the crawl, or says it in words nobody searches with.
-
It was retrieved but scored below the line
It is there, dashed, a little under the threshold. The fix is to make that passage more specific, or to add a verified answer for the question.
-
It was used and still misread
It is above the line, and the answer still gets it wrong. This is the rare case that is truly the model's fault, and it is fixed with a verified answer or a rule.
Without the trace, all three look exactly the same.
What shoppers see
Shoppers never see scores. They see a small “Sourced from this store” line under an answer that was built on your own content, and nothing under one that was not. An answer the assistant could not support is declined, with an offer of a person, rather than invented.
From diagnosis to fix
Every thumbs-down, every declined question and every request for a person lands in your review queue with its trace attached, together with anything the shopper typed about what was wrong. Fix it once with a verified answer, and the next shopper who asks gets the right reply.
For the people who check
If you are the person who tunes the assistant, the trace is also the fastest way to learn how your own content behaves: which pages carry most answers, which never appear, and which questions sit closest to the line.
Questions about the trace
What it records, and what you can tell from it.
What does a trace actually contain?
Every candidate the retrieval step considered for that answer, each with its relevance score, and which of them were selected as context. So you can see not only what the assistant used but what it nearly used and rejected.
Why is that more useful than the answer?
Because a wrong recommendation has several possible causes and they need different fixes. If nothing relevant was retrieved, the product or policy is missing or unindexed. If the right product was retrieved and scored below the cut, it is a tuning problem. If the right product was used and the answer is still wrong, it is a wording problem. The answer alone cannot distinguish those.
Can I tell whether an answer was grounded?
Yes, and it is recorded rather than inferred. Each answer is marked according to whether any retrieved content was actually used, and ungrounded answers collect in one place so you can see what your content does not cover.
Does keeping traces cost me anything?
No extra charge: they are part of every answer. They follow the same retention setting as the conversations they belong to.
See it on your own catalogue
Connect your store, ask what your shoppers ask, and read the trace behind the first answer. No card, and the Trial plan is free and does not expire.