Upload and extraction
A document arrives as a PDF, DOCX, text file or a URL. The first job is turning it into clean text — and this is where naive implementations lose most of their accuracy, because a pricing table flattened into a run-on line is worse than useless.
Waboxa converts documents to structured Markdown so tables stay tables and scanned pages are read rather than skipped, with a local extractor as a fallback when that service is unavailable.
Chunking and indexing
The text is split into overlapping passages and each is embedded into a vector index. The overlap matters: a policy that spans a paragraph break should not be findable only by half of itself.
Every stored passage carries the workspace and the business it belongs to, and every search filters on both. Isolation is a property of the query, not a promise in a settings page.
Retrieval, before generation
When a customer asks something, the question is embedded and the closest passages are retrieved first. Only then is the model asked to write, with those passages in front of it.
The order is the whole point. A model asked to answer and then check itself will confabulate; a model handed the relevant paragraph and asked to answer from it will mostly quote it.
The citation
The documents a reply drew on are written onto the message at the moment it is sent, and shown under the reply in the inbox.
They are recorded rather than recomputed later on purpose. Retrieval depends on the index as it stood at that moment, so re-running the search after a document was edited would answer a different question than the one the customer actually got.
This is what makes a wrong answer fixable. Without the citation, 'the AI said something incorrect' is an unanswerable complaint. With it, you know which file to open.
When to use a curated answer instead
Retrieval is the right tool for open questions. For the twenty things customers ask constantly, a human-approved answer is better — faster, exact, and incapable of inventing a price.
Sending those as a tappable list has a second benefit: the tap tells you which question the customer chose, from a list, inside their own conversation. That is a stronger intent signal than any page view.