Most explanations of how an AI chatbot works describe the pipeline in the abstract. Here's the same pipeline, followed from the point of view of one ordinary document: a two-page file called Refund-Policy.pdf.
Upload
A teammate drags the file into a document source. Nothing about answering questions happens yet — this is just an upload, the same as attaching a file to an email. The document is stored and queued for processing.
Chunking, minutes later
The file gets split into pieces — a paragraph about the 30-day window, a paragraph about non-refundable add-ons, a short section on how to actually request one. Each piece is small enough to be specific and large enough to still make sense standing alone, following the natural section breaks in the document rather than an arbitrary character count. This is chunking, and it's the step that decides how findable each part of the document will be later.
Embedding, seconds after that
Each chunk gets converted into a vector — a list of numbers representing its meaning, not its exact wording. The 30-day-window paragraph and the request-process paragraph end up as two separate vectors, positioned differently in that space, even though they came from the same file. Nothing about this step involves a customer or a question yet — it happens purely because the document was connected, in the background, before anyone asks anything.
Two weeks of nothing
The document sits indexed and unused. This isn't wasted time — it's exactly the state a well-prepared document is supposed to be in: ready, waiting, costing nothing extra to keep around, indistinguishable from every other chunk in the index until the right question shows up.
The question arrives
A visitor types: "can I get my money back if I cancel this month?" That question gets embedded too, and compared against every indexed chunk scoped to this organization's own content. The 30-day-window paragraph — written weeks earlier by someone who never saw this exact question coming — turns out to be the closest match by a wide margin.
Cited, in under half a second
The retrieved paragraph gets handed to the model along with the visitor's question, and an answer streams back citing Refund-Policy.pdf as its source — one stage in the full request timeline, most of which this document was never involved in at all. It just had to be sitting there, chunked and embedded correctly, when its moment came.
Six months later, an update
The refund window changes to 45 days. The file gets edited, auto-sync picks up the change, and only the affected chunk gets reprocessed — not the whole document, not the whole index. The next visitor who asks gets the new number, and the document's day starts again.
The takeaway
Nothing about this document's journey involved intelligence or cleverness at the moment it mattered most — the actual "smart" part happened weeks earlier, invisibly, when it was chunked well enough to be found. By the time a real question showed up, the hard work was already done.
Related reading
- Chunking and embeddings 101 — the step that determines whether a document like this one is findable at all.
- How vector search actually finds the right answer — the moment this document got matched to a real question.
- Auto-sync so your chatbot never goes stale — what happens to this document the next time it changes.