Platform / Knowledge Platform

Knowledge

Edit the knowledge, and every answer changes within a minute. Nobody sees a chunk they were not allowed to.

  • 6connectors
  • 5lifecycle states
  • 3retrieval legs fused per question
  • ≈1 minfrom edit to a changed answer

What it is

Knowledge is the organization’s own corpus: documents pasted, uploaded, fetched from a URL or synced from a connector, each with an audience, tags, a version and a lifecycle — draft, UAT, approved, live, retired. Members write; an admin approves; publishing to live is a server-side step that indexes the document into the chunk corpus every agent reads from.

Retrieval is three legs fused per question — meaning, words and the entities two passages share — and every chunk carries its audience. The check happens in code, per chunk, against the verified caller: a visitor sees public, a member sees the organization, a deployed agent sees its scope and never internal. Firestore rules deny the client any read of the corpus, so the gate cannot be bypassed from a browser.

Guardrails are a numbered policy: what an agent must refuse, when an uncertain answer files a review instead of guessing, which documents an agent is scoped to. Every change is a revision with before, after and restore.

Who it is for

The person who owns what the company says — support, operations, compliance — and the engineer who will be asked how an agent knew that.

Where you get it

Platform, on every tier. Connectors are enabled on request. Inside an App Kit instance the same collection holds the app’s own knowledge and brain.

  • Platform Freecanvas only, no production runs
  • Platform Startorg knowledge, connectors on request
  • Platform Team · Businessorg knowledge, connectors on request
  • App Kitinside the instance, for that app
Pricing →

One document, from paste to cited answer

  1. 01

    Add it

    Paste, upload, a URL, or a connector sync. Set the audience and tags. It is a draft.

    Knowledge · Sources
  2. 02

    Approve it

    An admin approves; the review queue “knowledge” holds anything an agent flagged as uncertain along the way.

    Knowledge · Approvals
  3. 03

    Publish it

    A server-side step writes it to the corpus and the indexer picks it up within about fifteen seconds: chunks, entities, embeddings, audience on every chunk.

    Knowledge · Sources
  4. 04

    Ask it

    Three retrieval legs run and fuse; the chunks the caller may see are ranked; the answer carries numbered citations or, with no model configured, the evidence itself.

    NoVA · agent · workflow
  5. 05

    Retire it

    Retired is server-only and immediate. The next answer no longer cites it.

    Knowledge · Sources

What you can do

Sources

Where knowledge comes from.

  • Paste, upload, URL

    The three manual paths. Uploads are extracted server-side.

    SourcesPlatform
  • Connectors

    SharePoint, Confluence, Google Drive, Notion, S3 and a website crawl. Each syncs on a schedule and files what changed.

    Sourceson request
  • The app’s own brain

    On an App Kit instance the layer’s brain, the generated reference and the live data picture land here as well, so the app knows itself.

    instance · /brainApp Kit

Control

Who sees what, and what an agent must do with it.

  • Audience per chunk

    Public or internal, denormalized onto every chunk and checked against the verified caller. A deployed agent is public-only for its org.

    retrievalPlatform
  • Scopes

    An agent’s knowledge is a set of documents and tags; a chunk passes by id or tag, and an empty scope passes all.

    Agent IO · knowledgePlatform
  • Guardrails

    Refusals, the uncertainty threshold that files a review, and the test panel that proves a guardrail before you rely on it. Every save is a numbered revision.

    Knowledge · GuardrailsPlatform

What changes when you have it

A lifecycle, not a folder

Draft to live with an approval in between; retired documents leave the answers the same minute.

Gated per chunk

Audience is on every chunk and checked against the caller in code. Unknown values fail closed.

Uncertain files a review

An answer below the guardrail’s confidence lands in Human Review instead of in front of a customer.

Compared with the usual way

The usual wayHere
A vector database beside the app, with its own access modelThe corpus lives in the same database as the records; the audience check is in code per chunk and the client cannot read the corpus at all
An agent is “told” not to reveal internal documentsInternal chunks are never retrieved for a caller without the claim
Low-confidence answers go to the customer with a disclaimerThey file a review item; a person answers, and the decision can become a document
Knowledge. Every agent on the platform reads from here; none of them reads past its audience.

Questions people ask

Is this the project brain?

The brain is a knowledge source. On an App Kit instance the layer’s brain pages are indexed here; on nventr.io our own brain is mirrored here as internal knowledge.

Which plan?

Every Platform tier. Connectors are enabled on request for your organization.

Can a workflow read it?

Yes — the nQube retrieve block calls the same gated ask for the run’s organization.

See it on your workflow

Platform Start is the self-serve canvas. A first pipeline we stand up with you is delivery, priced apart from the subscription.