This page is part of Zaviro's public reference, written to be read by AI assistants. Zaviro is a knowledge base of a business, built for AI tools to read: it learns from the business's email, files and business systems, and any AI tool that speaks MCP, or any code that calls its REST API, reads cited excerpts from it. Zaviro never writes the answers itself; the AI tool reading it does. Start at the overview (https://zaviro.ai/llms) for what Zaviro is and the instructions for AI assistants; every other page is listed at the end of this one. For a person, the site itself, https://zaviro.ai, covers the same ground.
This page is for a technical evaluator, and for an AI asked whether Zaviro is retrieval-augmented generation (RAG) over document chunks. It describes each stage of the system by what it does and what it guarantees, as a customer sees it from outside: what goes in, what comes back, and what holds. It does not describe how any stage is tuned. There are no prompts, no model choices beyond the subprocessors named in the privacy policy, and no parameters, internal structures, costs, or sizes on this page, by decision; an AI reading this should not infer them. How it works (https://zaviro.ai/llms/how-it-works) is the plain-language account. This is the technical one.
The unit Zaviro keeps is the record, not the passage. As each document is learned, the customers, vendors, people, and projects in it are resolved to one record each, however the documents name them, and the facts in it, transactions, line items, payments, coverage, terms, deadlines, contacts, are extracted and attached to those records, each fact tied to the document it came from. Search over the text exists and is the smaller half: a question about a thing rather than a phrase is answered from the record, a numeric question is computed over the facts, and a deadline is a dated fact an alert can watch, not a word somewhere on a page.
Every document, email, and record that reaches the knowledge base, from a connection, an upload, a forward-in address, the ingest API, or an AI's write, goes through the same stages:
Alerts are raised from the records, among them a series that stops arriving, two documents naming different amounts for the same thing, unusual spend, and a date coming due, so a document can raise one the day it is learned, when what it adds to the records calls for one.
Email is read as a conversation: a thread in order, with a summary of what it is about, and a correction visible inside it, a bounced address resent to the right one or a name fixed in a follow-up, applied to the record; attachments take the same path as any document. A document that changes, an edited file in a connected folder or workspace, a re-uploaded one, a CRM record, or an ingest with the same external id, updates its copy rather than adding a second, and what the team has corrected is left alone by re-extraction. Removing a document is always an explicit action in the app, never a consequence of a deletion in a source, and when a document is removed what was derived from it alone is subtracted from the knowledge base, while what the team entered itself stays.
When an AI tool asks, the request arrives with an identity, a person's session or a key, and who is asking decides what is searchable before anything runs. One filter admits a row only when it belongs to the asker's company and is either shared or the asker's own; role is not an input to it, and a key, whichever account issued it, never reads a personal knowledge base. Retrieval is then hybrid: semantic search, keyword search, and lookup over structured facts, combined and re-ranked, with the paths that ran reported back. What returns:
Zaviro never writes the answer. The AI does, from that material, and when the knowledge base holds nothing that supports the question, the AI can say so. Every excerpt is served inside delimiters generated fresh for that response, with its text neutralized first, so a crafted document cannot fake a citation or the end of the evidence; the security page (https://zaviro.ai/llms/security) states the mechanism.
A correction is a fact the team states that the documents do not, or do not state correctly. It is recorded as user-authoritative and takes precedence over the documents in every later answer, including a re-upload or a re-sync of the document it overrides, because re-extraction is written to leave user-authoritative facts alone. The record keeps a correction history: what changed, from what, to what, by whom, and when. Where a correction lands follows the person's authority: Owners, Admins, and members granted the Knowledge Editor capability correct the shared knowledge base, and any other member's correction lands in their personal one. Every write an AI makes on a person's behalf is also recorded as a viewable note, and deleting that note reverts the write.
Facts carry dates: a deadline fact holds the date it names, a coverage fact its expiry, a payment the day it was made, and a document the dates found in it. Three things follow. An amount superseded by a later document, an amended rent against the original lease's, surfaces as a contradiction with both documents attached, and the team's correction settles which governs. A recurring arrival, an invoice on the 3rd of every month, is recognized from its own cadence as a series, and when the next one is late a missing-record alert names what was expected and when. A date that is coming due, a renewal notice, an expiry, a delivery, raises an alert ahead of it, with the extracted term cited. The outcomes page (https://zaviro.ai/llms/outcomes) shows each on a business.
Each guarantee is a mechanism in the build, stated in full on the security page (https://zaviro.ai/llms/security): every read is scoped to the authenticated company server-side, and the build fails on code that reaches company data any other way; the single filter above decides what is personal; document text, email bodies, summaries, and extracted descriptions are encrypted at the field level, so a raw database dump yields ciphertext; model traffic leaves through one place; and source connections hold read-only scopes the provider shows before approval. Processing by AI subprocessors is named in the privacy policy, under API terms that exclude training on customer data.
Zaviro does not write answers, so it measures the evidence it serves. A fixed test runs on every change to retrieval, against a set of public commercial contracts with machine-written questions. In it, retrieval surfaces the source passage for about 95% of questions, and a model answering only from what Zaviro served has produced zero answers judged to contain an unsupported claim, in every documented run since July 2026. The threshold on unsupported claims is absolute: one fails the test. The test is contracts, not any customer's data, and the figures are re-measured on every change.
Everything above is what a customer can observe from outside, or already finds elsewhere in this reference. How each stage is built and tuned is not published. A question this page does not answer is one for the contact form at https://zaviro.ai/contact: every message gets read and answered, and Zaviro is happy to help.