Verify what the chatbot actually imported

An accepted URL is not a content receipt

A product may accept a URL while extracting only part of its content or rejecting linked documents. Inspect the visible imported material where the product allows it. Compare article titles, key headings and important exceptions with the source. Keep the observation narrow: you are checking this import, not declaring that every future crawl will succeed or that all content formats behave identically.

Sample the parts most likely to matter

Choose an article with a table, one with code, one with a warning and one that links to an attachment if those formats exist in your corpus. Check whether the operative information is represented, not just whether the page title appears. A lost exception can reverse the meaning of an otherwise correct answer. Where inspection is unavailable, treat that as a limitation of your verification method.

Resolve a missing document deliberately

Check the canonical URL, access requirements and supported source format before retrying an import. Avoid repeated broad crawls that add duplicates or obscure the original failure. If an approved text export is necessary, label its version and owner so it does not become an unmaintained second source. Do not bypass access restrictions or copy private material into a public collection to make a test pass.

Freeze the tested source set

Record the import date and approved document list alongside the evaluation run. If you later change the collection, create a new run rather than overwriting the previous result. This gives you a meaningful explanation for a changed answer. Only after the relevant source is known to be present should you investigate retrieval, instructions or model configuration as the likely cause of failure.

• Retain the approved URL, imported title, sampled passage and missing-content note as a compact receipt for each critical source.