Second check in chat: how prices, deadlines and promises are held against the source
How XICBOT holds prices, deadlines and promises against the source material after writing, what counts as evidence and where the check ends.
In plain language
An assistant can name a figure that is in none of your sources. So XICBOT checks the answers of the assistant once more. It looks for prices, time limits, amounts and promises such as free of charge. Then it checks if they are in your sources.
If one is missing, it is usually taken out of the text. In its place the text says: not backed. Most of the time the visitor also gets a short note.
The check only sees if a figure is in a source, not if it is right.
An assistant that answers from your sources can still state a detail that is not in them: a delivery time from the general knowledge of the language model, a percentage that belonged to a similar question, or a “free of charge” that nobody promised. In sales weeks, when a shop states a particularly large number of prices, deadlines and conditions, such an error carries real weight. XICBOT therefore holds hard claims against the source material once more after the model has written, as a rule replaces anything uncovered with a placeholder and makes the correction visible. This article explains what counts as a hard claim and what counts as evidence, why the comparison is deliberately lenient, which results an answer can carry and where the check ends.
Labels in the illustration
- Model draft
- Check against the excerpts
- Delivered answer
- The Alpin 30 costs €89.90,
- arrives in 2–3 days,
- and shipping is free of charge.
- The Alpin 30 costs €89.90,
- arrives in 2–3 days,
- and shipping is
- [not backed].
- Excerpts for the question
- [1] Product data: price €89.90
- [2] Shipping page: 2–3 days
- Price
- €89.90
- found in excerpt [1]
- Period
- 2–3 days
- found in excerpt [2]
- Promise
- free of charge
- in no excerpt
- Note (default setting)
- The stored sources do not
- back a promise.
- Verification: Uncovered claims
- found in an excerpt
- in no excerpt
- [not backed] replaces the uncovered part of the text
- Limit
- No source material (no hit, no free-text block, no tool): nothing is replaced
- Answer without a used citation and without a tool: entry in the gap list
Why answers are checked again after the model
The first safeguard of an assistant bound to sources comes before writing: for every question it receives numbered excerpts from your documents and is meant to answer from them alone. How this binding works is described in the article on hallucinations and the knowledge base. It lowers the risk of invented details but does not rule it out. A language model assembles text word by word, and a figure can emerge along the way that is in none of the excerpts. The German Federal Office for Information Security (BSI) describes this behaviour in its paper “Generative AI Models – Opportunities and Risks for Industry and Authorities” under risk R4 “Lack of Output Quality”. The starting point there are models without access to real-time data: many of them nevertheless process inputs on current topics and “accordingly invent content when generating outputs that was neither part of the input nor of the training data”. For language models the BSI also describes such hallucinations as a significant issue, because the outputs often appear convincing.
For a shop this is not a theoretical question. During a sales week the site lists sale prices, delivery times, conditions for free shipping and notes on returns, often with exceptions for reduced goods. When a visitor asks about them, the answer states exactly these details, and these are exactly the details the visitor bases a decision on. A wrong delivery time or a promised free delivery that does not exist comes back later as a complaint. This is why XICBOT runs a second check after the model, holding hard claims from the draft against the material that counts as evidence. The drawing above shows the sequence with an example: price and delivery time are found in the excerpts, the promise “free of charge” is not.
In the case of potentially critical implications, outputs from generative AI models should be reviewed, cross-referenced with information from additional sources if necessary, and, if required, finalised through manual post-processing before further use.
Measure M20 names three steps: review, cross-reference with additional sources if necessary and, if required, finalise through manual post-processing. It particularly highlights outputs used for purposes with external impact and gives content for a company’s internet presence as an example. A chat answer to customers has an external impact, yet in a live conversation nobody can read every answer before it goes out. The automatic second check therefore takes over the comparison for a narrowly defined kind of detail, and it does so in every finished answer from the model. Manual post-processing moves to before and after: into maintaining the sources before a sale and into reviewing the reported answers afterwards.
What counts as a hard claim
The second check searches the draft for details that can be compared with a source without interpretation. These include sums of money with a currency, percentages, deadlines and periods such as days, weeks or months, quantities with a unit symbol such as kg or cm, counts of pieces and people, e-mail addresses, web addresses and phone numbers. On top of that comes a short list of promise words: “free of charge”, “at no cost”, “anytime”, “lifetime”, “unlimited” or “guarantee” turn a piece of information into a promise. For these words the stem is compared, so “guaranteed delivery” in the draft is covered by “guarantee” in the source. How this plays out for environmental claims with a figure, such as a share in percent, is described in the article on environmental claims in chat.
Just as important is what is not on the list. The check works with fixed patterns and recognises only what matches one of them. Anything that cannot be compared unambiguously is deliberately left unchecked, because an uncertain comparison would produce more false alarms than benefit. As a rule, the following are not covered:
- Product names, colours, features and other properties without a figure.
- Numbers written as words: “three days” is not recognised, “3 days” is.
- Calendar dates, such as the end of a sale: a date is not a period in the sense of the patterns.
- Promises hidden inside a compound word, such as the German “versandkostenfrei” (free shipping), as long as the promise word does not stand on its own.
- Judgements and advertising words such as “cheap”, “fast” or “sustainable”.
What counts as evidence
The material a claim is held against decides the result. For this purpose XICBOT combines the excerpts, the tool results, the visitor’s messages and the role, topic scope, handover rule and greeting of the version. Maintained facts and structured master data such as opening hours, locations, contacts or price ranges are also available to the model, but they are not part of this material. If an answer states a price or a phone number taken only from these fields, the second check therefore usually replaces it with the placeholder. What the assistant itself said in earlier answers is explicitly excluded. Otherwise an invented figure from the third answer could cover the same figure in the fifth, and one invention would serve as evidence for the next.
| Material | Counts as evidence | Counts as source material |
|---|---|---|
| Excerpts from the knowledge base for the current question and free-text blocks | yes | yes |
| Tool results, such as product data or the cart | yes | yes |
| The visitor’s current message | yes | no |
| The visitor’s earlier messages in the same conversation | yes | no |
| Settings of the version: role, topic scope, handover rule, greeting | yes | no |
| Maintained facts and master data fields of the version | no | no |
| Earlier answers of the assistant | no | no |
The right-hand column explains a distinction that becomes important later. Only the excerpts and the tool results count as source material in the narrower sense. Only if some of it was available is “not covered” a statement about the answer. The excerpts also include free-text blocks, short text sources that the model receives with every question, whether the search hits them or not. The visitor’s messages count as evidence because an answer may pick up what the visitor wrote: if they state a quantity or a period, the answer may repeat it without a placeholder appearing.
A lenient comparison: covered does not mean correct
The comparison is a text measurement, not a second trip to the language model. A model that checks its own answer costs another call and returns the same kind of guess that is supposed to be checked. Instead, every hard claim is brought into a comparison form and searched for in the material: for figures only the digits, without thousands separators, spaces and zero decimals, so that “1,299.00” and “1299” are the same figure; for promises the word stem; for addresses the spelling regardless of upper and lower case. A claim counts as found if its comparison form occurs anywhere in the material.
This direction is intended. Reporting a covered figure as invented would be worse than not checking at all: the visitor would see placeholders where nothing was wrong, and the operator would lose trust in the reports. Missing an invented figure, on the other hand, costs that one case. This leads to a clear limit: what gets reported is anything that appears nowhere in the material, not anything that is wrong. Whether a figure stands in the right context is not something the check measures. We ran the examples in the table through the check.
| Draft | In the material | Result |
|---|---|---|
| Delivery in 2–3 days | Delivery time 2–3 days | covered and correct |
| Returns within 14 days | Returns within 30 days; orders placed by 14:00 leave the warehouse the same day | covered although wrong: the 14 is in the material, only with a different meaning |
| Exchange within 3 days | Delivery time 2–3 days | covered: a single-digit figure appears in almost any material |
| Returns within 60 days | Returns within 30 days | not covered: placeholder instead of the period |
| Returns possible anytime | no “anytime” in the excerpts | not covered: placeholder instead of the promise |
Covered is not a seal of approval
The result “All claims covered” means: every recognised hard claim occurs in the material. It does not mean the answer is right. The second check is a net for figures and promises that come out of nowhere, not a fact check.
Five results on every answer
The outcome of the second check is stored with every answer and shown in the conversation view below the assistant’s answer, labelled “Verification”. Five results are possible. Two of them sound similar and still mean different things: an answer without a checkable claim was measured and is unremarkable, an answer that was not measured did not go through the check at all. This way a missing result is not booked as a good result.
| Result | Meaning | Effect on the text |
|---|---|---|
| All claims covered | Hard claims found, each one found again in the material | none |
| No checkable claim | Measured, but nothing found that could be compared, for example a greeting or a follow-up question | none |
| Uncovered claims | At least one claim is in no material, although excerpts or tool results were available | placeholder at that point, correction sent to the visitor |
| Answered without source material | Hard claims found, but neither excerpts, including free-text blocks, nor tool results were available | no replacement; the point is marked in the chat window, and in the default setting a note also appears below the answer |
| Not measured | No finished answer from the model: aborted, emergency reply, blocked request or curated question-answer pair | none |
For uncovered claims, the view also states their total number and lists the points individually, up to eight per answer, with their type, such as price, deadline or promise, and their position in the answer text. Only type and position are stored for this, not the wording a second time: the wording is already in the answer itself, and when the answer disappears at the end of the retention period, the pointer disappears with it. For a curated question-answer pair, “Not measured” is not a flaw. It goes out in the wording you wrote, and there the text itself is the source. All the more important is that a pair containing a price or a deadline is updated after a sale, because no second check catches an outdated detail in it.
Placeholder and correction for the visitor
The answer appears in the chat window piece by piece as it is written. The second check only runs once it is complete, so that it does not delay the start of the answer. This has a consequence that should be named openly: the visitor may have read the uncovered claim for a moment before it is replaced. This is why the correction does not happen silently. It goes to the chat window as an event of its own, with the full corrected text, and the speech bubble then shows that text. A bubble that silently differs from the stored answer would be the bigger trap.
- The uncovered claim is replaced by the placeholder “[nicht belegt]” in German conversations and by “[not backed]” in conversations in any other language; the chat window also marks it.
- The rest of the text stays as it was, including the covered claims and the source references below the answer.
- If a product card has split the answer across several speech bubbles, each bubble is usually corrected on its own.
- In the default setting, where the assistant answers only from your sources, a short note follows stating which kind of detail is not backed and asking the visitor to follow up if needed.
- Handover to a person remains available in the conversation.
The note names the type of detail, for example a price, a deadline or a promise, and says that the point was replaced by the placeholder. It does not repeat the original value. The same principle applies to product cards: the result of a tool goes into the chat window as a card and to the model as data, and prices, quantities or deadlines in the sentence next to the card are held against these data. How product cards work in the chat is described on the page about product cards.
When there was no source material
If there are neither search hits nor free-text blocks nor tool results for a question, there is nothing a claim could be held against. A separate rule applies to this case: the answer text is left untouched, nothing is replaced, and the result is “Answered without source material”. The chat window also marks the point, and in the default setting a short note below the answer says that the detail is not backed by the stored sources, asking the visitor to check before relying on it. “Not covered” would be a statement about an empty shelf here, not about the answer. An assistant whose knowledge base contains nothing on a topic would otherwise deliver nothing but placeholders, and the result “Uncovered claims” would lose its meaning. Stored free-text blocks, by contrast, count as source material even without a matching hit: if there are any, an uncovered claim is replaced even when the search found nothing for the question.
As a rule, an assistant bound to sources says anyway that nothing is stored on the subject when no excerpt carries the answer. This refusal is a prescribed behaviour, is counted as a refusal and cites no evidence; what a helpful refusal sounds like is described in the article on declining instead of guessing. If “Answered without source material” still appears on an answer containing a price or a deadline, that is a signal: either the binding to sources was relaxed for this version, or the matching section is simply missing for the question. Both can be fixed, the first in the settings of the version, the second in the knowledge base.
Reviewing uncovered claims and gaps
Besides the second check there is another measurement that is easily confused with it. The second check asks per claim: is this figure or this promise in the material? Gap detection asks per answer: did the assistant use an excerpt at all? A finished answer in which not a single citation was used is recorded as a gap, as is a question for which the assistant found nothing. Excluded are very short inputs such as a greeting, answers that rely on a tool, and assistants without any stored knowledge. An answer can pass both checks, only one or neither.
| Question | Second check | Gap detection |
|---|---|---|
| What is looked at? | every hard claim in the answer | the answer as a whole |
| When does it trigger? | a claim is in no material | no citation and no tool used, or nothing found |
| What does the visitor see? | placeholder, marking and, in the default setting, a note | the refusal or an answer without a source reference |
| Where do you see it? | result on the answer, switch in the conversation view | gap list in the conversation analytics, from the Pro package |
After a sales week this leads to a short routine. In the conversation view, the switch “Only answers with uncovered claims” shows the conversations in which something was replaced. Each point raises the same question: is the detail missing from the source although it is correct, or did the model add it? The gap list in the conversation analytics shows, from the Pro package, the questions for which no source carried the answer. How both lists turn into regular maintenance is described in the article on keeping the knowledge base up to date.
- Filter conversations with uncovered claims and read the replaced points one after the other.
- For each point, decide whether the detail is missing from the source or was added.
- Add missing details to the knowledge base, sale prices and deadlines there rather than in master data fields or maintained facts, which are not part of the comparison; create a curated question-answer pair for fixed information.
- If a price or a deadline changes after the sale, update the source and the affected pairs on the same day.
- Then ask the same test questions again and compare the result on the answer.
Before the sales week: the second check in a test conversation
Whether the second check takes effect in your assistant can be shown before a sale with a few test questions. Useful are questions whose answer could contain a hard claim that is missing from the sources, and counter-checks where everything is stored. The promise word should not be part of the question: anyone asking “Is shipping free of charge?” supplies the words “free of charge” themselves, and they count as covered. Ask the questions in the test conversation and read in the conversation view which result is shown on the answer. How to set up repeatable test cases is described in the article on acceptance testing with test cases.
| Test question | Situation in the sources | Expected result |
|---|---|---|
| How long does delivery take? | delivery time stored | All claims covered, answer with citation |
| What does shipping cost during the sales week? | nothing stored about free shipping | a placeholder if the draft promises “free of charge” |
| How long can I return reduced goods? | only the general return period | the general period with citation; a different figure is usually replaced |
| Until when does the sale price apply? | sale period stored | period from the source; the check does not cover a calendar date that is mentioned |
| Hello, is anyone there? | not needed | No checkable claim |
What the second check does not do
The second check is one of several safeguards, and it is deliberately narrow. It replaces neither maintained sources nor the binding of answers to those sources, and it does not make an answer legally safe. What to observe for prices in chat is described in the article on prices in chat; in an individual case, professional advice remains the right route. An assistant can still err with the second check in place. The binding to your sources and the cited source reference are described on the page about the knowledge base.
- It recognises the uncovered, not the wrong: a figure that appears in the material with a different meaning counts as covered.
- It sees only fixed patterns: numbers written as words, calendar dates, properties without a figure and promises inside compound words stay untouched.
- It counts the visitor’s message as evidence: a wrong figure brought along stays in place if the assistant repeats it.
- It does not hold claims against maintained facts and master data fields: a correct detail taken only from these fields is usually replaced.
- It does not run over curated question-answer pairs, whose wording you are responsible for.
- Without search hits, free-text blocks and tool results it replaces nothing; the claim stays, marked and in the default setting with a note below it.
- It comes after the answer: the visitor may have seen the original value briefly, hence the visible correction.
Frequent questions
What exactly does the second check look at?
It searches the model’s finished answer for hard claims: sums of money, percentages, deadlines and periods, quantities with a unit, contact details and promise words such as “free of charge” or “anytime”. Every claim found is searched for in the material that counts as evidence: excerpts, tool results, the visitor’s messages and some settings of the version. As a rule, it does not cover properties without a figure, numbers written as words or calendar dates.
Does “covered” mean the claim is correct?
Not necessarily. Covered means the claim occurs in the material. If the same figure appears there with a different meaning, it also counts as covered. The check finds claims without an origin; it does not check context. For that, maintained sources and test questions are usually the more reliable route.
Does the visitor see that something was replaced?
As a rule, yes. The replaced point carries the placeholder “[not backed]” and is marked in the chat window, and the correction reaches the chat window as an event of its own. In the default setting a short note with the type of detail follows as well. Because the check runs after the answer is finished, the visitor may have seen the original value briefly.
Why is nothing replaced without source material?
Without search hits, without free-text blocks and without tool results there is nothing a claim could be held against. “Not covered” would then be a statement about an empty knowledge base, not about the answer. Nothing is replaced for that reason; the chat window marks the point, and in the default setting a short note appears below the answer. The result is “Answered without source material”, and it usually shows that a section is missing for the question. If free-text blocks are stored, they count as source material, and replacement happens even without a matching hit.
Are curated question-answer pairs checked as well?
No. A curated pair goes out in the wording you wrote and carries the result “Not measured”, because there the text itself is the source. Prices and deadlines in pairs should therefore be updated after every change, for example at the end of a sale.
Does the second check need another call to the language model?
No, it is a text comparison against the answer’s material. It runs once the answer is finished and therefore does not delay the start of the answer in the chat window. As a rule, the visitor only notices it when something is replaced or marked as not backed.
Ask it yourself, Assistant of this website
Automated assistant