What the number under an answer is actually pointing at

The answer is right, or close to right, and the numbered source under it opens a page that does not obviously contain the line you just read.

In rough order of how often each one turns out to be the answer. Work down rather than across: each carries a way to tell whether it is yours before you change anything.

• 1

  The link is the document address, not the passage address

  Why

  Your material is split into passages for matching, but a passage carries no address of its own. The label and the link on a numbered source are the title and the address held against the whole document, which for a crawled page is the page title element and the page address as fetched. A long page therefore always opens at the top, however far down the useful sentence lives.

  How to confirm it is this one

  Open the linked page and search it for a distinctive phrase from the answer. Finding it, somewhere well down the page, is this and nothing else. Not finding it at all is one of the causes below.

  Fix

  Nothing to configure, because there is nowhere for an anchor to be stored. The practical answer is shorter pages under their own headings, which improves the matching at the same time as it improves the link.

• 2

  The number was written by the assistant, and it is taken at its word

  Why

  The assistant is asked to put a number in square brackets straight after the claim it supports. The finished reply is then scanned for numbers in square brackets, and any that fall within the range of passages it was actually shown become the sources listed underneath. A number attached to the wrong sentence survives that unchanged, because there is no second check that the passage says what the sentence says.

  How to confirm it is this one

  The markers are visible in the reply itself, drawn as small raised numbers inside the text. Count the distinct markers in the reply and the chips underneath: they match one for one, always, because the chips are built from the markers.

  Fix

  This is not a setting. It gets worse with more passages in front of the assistant at once, so bring Passages it reads on the Behaviour tab back towards its default of six if you have raised it.

• 3

  Every page on the site carries the same title, so every source reads alike

  Why

  The label on a crawled document is the title element of the page as it was delivered. A site whose title is the business name on every page produces sources that are all labelled identically, and the only thing separating them is the address behind each one.

  How to confirm it is this one

  Open two different numbered sources under the same answer. Identical labels with different addresses is this, and it means the citation was probably correct all along and merely unreadable.

  Fix

  Give the pages distinct titles and re-index the source. Nothing about the assistant changes; what changes is your ability to tell one source from another at a glance.

• 4

  The reply came out of the cache, and the sources came with it

  Why

  An answer that cited something is kept for that assistant and returned again when a later question is close enough in meaning. The sources stored beside it are returned exactly as they were, so they were chosen for the question somebody asked first rather than for the one you just typed.

  How to confirm it is this one

  In the browser network tab, the response to the chat request opens with a meta event carrying cached set to true. Asking the same thing in genuinely different words misses the cache and produces sources picked for your wording.

  Fix

  Nothing, and nothing worth doing. The store is cleared whenever a source finishes indexing new text and expires within a day, and it is the reason a repeated question costs nothing.

• 5

  The last source is one the assistant only saw the beginning of

  Why

  The passages are assembled within a fixed budget, and when the final one does not fit it is cut short rather than left out. That is deliberate, so a number can never point at something the assistant was never shown, and the price of it is that the highest numbered source under a long answer can be a passage that arrived half read.

  How to confirm it is this one

  It is always the last chip, never the first, and only on assistants reading a lot of passages. Lower Passages it reads and ask the same question again. If the odd source stops appearing, it was the tail of the budget.

  Fix

  Leave Passages it reads at six unless you have a reason. Raising it past the point where the budget fills adds no material, it only means the final source arrives truncated.

A source with no link at all is not a fault. A document carries an address only when it came from a crawl, so anything you uploaded as a file is labelled and numbered like everything else and is drawn as plain text rather than as something to click. If you want the citations on a subject to be clickable, that subject has to be on a page the crawler can reach rather than in a document on your desktop.

Can the link be made to jump to the right part of the page?

No. The address recorded against a document is the address the page was fetched from, with nothing appended, so there is nothing for a jump to be built out of. Splitting a long page into shorter ones is the only route to a link that lands somewhere useful.

Why does the same number appear twice, in the sentence and underneath?

They do different jobs. The one in the sentence says which claim is being supported, which is the part that matters when an answer makes several. The one underneath says what supported it and gives you somewhere to go.

Could it ever cite something the assistant was not shown?

No, and that is enforced rather than hoped for. The list of sources is built only from the passages that actually fitted in front of it, and a number outside that range is ignored rather than turned into a chip.