01 Deployed
The conversation itself.
Streaming answers, a model selector bound to your allowlist, a region selector, sessions that persist, attachments, markdown rendering and voice dictation.
The first capability your people touch, running the day the deployment is stood up.
Frontier chat over your own documents, limits stated.
Chat is the first capability a deployment gives you, not the product it is. Streaming answers over your own allowlist, an optional knowledge base over your own documents, and usage each person can read for themselves. It runs under your content policy, on the day the environment lands.
A working surface, not a demo.
THE SURFACE · WHAT STAFF GET
Every deployment has general chat from the day it is stood up. It runs over the governed catalogue, under your content policy, metered against each person who uses it. The work people were already doing in browser tabs now happens somewhere you can answer for.
02 Deployed
Your documents, toggled on.
Upload documents, watch ingestion status, and turn the knowledge base on for a conversation. Answers come back with citations to the source document.
03 Deployed
Each person sees their own usage.
A quota gauge, tokens by day, storage used and document count. Tokens, never dollars: the product holds no currency, and your cloud invoice is the financial record.
Upload, index, ask, or decline.
KNOWLEDGE BASE · FOUR STEPS
- 01You
Upload
Documents land in object storage inside your own account, in your storage region, with their metadata and status listed in the product.
GatePDF only today. An unsupported format is turned away at the door rather than accepted and quietly ignored.Decides: the upload check - 02Automatic
Ingestion
Ingestion is event-driven into a vector index that sits in the same storage region. Progress is visible while it runs.
GateUntil ingestion finishes the document is not retrievable, and the list says so rather than pretending otherwise.Decides: ingestion status - 03Per conversation
Ask with the base on
The knowledge base is a toggle inside a conversation, so a person always knows whether an answer was drawn from your documents or from the model alone.
GateThe toggle is explicit. Retrieval is never an ambient behaviour you have to guess at.Decides: the person asking - 04The answer
Cited, or declined
An answer returns citations to the source documents. When retrieval finds nothing, the assistant retries with alternative phrasings and then declines rather than composing something plausible.
GateThe decline fires on zero retrieval rather than on a confidence score.Decides: the retrieval result
Where retrieval stops.
LIMITS · THE KNOWLEDGE BASE
Before you load anything
Everything in a deployment's knowledge base is readable by every user of that deployment. There are no per-user or per-document permissions on retrieval. Load only what everyone with an account may see. If your requirement is a permissioned document library, read platform or gateway before you scope anything.
| Limit | What it actually means | What to do about it |
|---|---|---|
| Retrieval is deployment-wide | No permission filter at query time, no permission labels at ingest | Load only what every user may see |
| Citations are returned, not validated | Nothing independently checks that the cited passage supports the sentence | Treat a citation as a pointer to check |
| PDF only, today | Other formats are not ingested into the knowledge base | Convert, or keep them out of scope |
| Decline fires on zero retrieval | No groundedness score and no confidence threshold exist | Assume a confident wrong answer remains possible |
| The invocation log holds no prompt text | Model, caller identity, latency and token counts only | Use it as usage evidence, never as a transcript archive |
Each row is what a buyer needs in order to design around it in week one rather than discover it in month three. Everything else we do not claim.
What the table above lets you plan.
DESIGNING AROUND IT · WEEK ONE
It's not a black box. You can see how it's set up, and that's exactly what lets you answer for it.Why the table above exists
Knowing where retrieval stops is what lets you shape the deployment around it: keep the sensitive intake out of the index, put a permissioned corpus behind the other product, and use chat for the work it is genuinely good at. The glass box is not a slogan about transparency. It is the reason your security owner can sign. How we apply it everywhere else.
Chat is the first capability.
POSITION · FIRST, NOT ONLY
Day one chat ends the shadow usage, and that alone is worth the deployment. It is not, however, what organizations end up using every day. That turns out to be governed business software: a scheduling tool, a roster with encrypted fields and an audited reveal path, a secure intake drop, a forecasting tool over their own records. Most of those never call a model at all, and we treat that as a strength rather than an embarrassment. Inference is not always the answer.
Where the day-to-day work actually lands.
Applications framed inside the same dashboard, each on its own origin, with credentials tagged to the person using them and an access record written before the credential exists.
org.tech · Modules→
Your way of doing things, productionized.
We sit inside the work, watch how it is really done, and build that into the environment as a capability. Scoped and priced per deliverable, never bundled into a platform fee.
org.tech · Engineering→
When the gateway fits better.
THE OTHER PRODUCT
If your people already use an AI assistant and the real problem is your document library, orgOS chat is probably not the thing to buy first. The operational knowledge gateway gives that assistant governed, scoped access to your own files, with server-enforced access control per person. It carries a different data-boundary trade: tool results travel to the assistant's vendor. Neither product is the right answer to every question, which is why there is a page for choosing.
- 01Platform or gatewayThe decision, with the cases where each one is the wrong choice.→
- 02Access and the data boundaryWho can read what, and exactly where data travels.→
- 03What the gateway does not claimIts limits, printed the same way these are.→
01Can we restrict a document to one team?
Not in the knowledge base today. There is no per-user or per-document permission on retrieval, so the control is what you choose to index. If per-person scoping is the requirement, the gateway is the product that has it.
02Which models can staff choose in chat?
Whatever your allowlist permits, from the generated catalogue, with the region selector bound to your configured endpoint regions. See model access.
03Can an administrator read our conversations?
Yes, and the read itself is written to the audit record as it happens. That is deliberate: the power exists, and so does the trace of using it. See administration.
04Can it connect to our document management system?
orgOS does not mirror an external library. The gateway does exactly that, one approved library at a time, read-only and revocable in one action. How it works.
05What stops chat being used for something it should not be?
Your content policy, attached to every call and enforced as a deny in cloud identity policy rather than an application setting. What the policy says is yours to set: denied-topic and personal-information rules are supported by the policy engine and configured per client.
⎯⎯ Book the strategic assessment ⎯⎯
Your private AI, inside your control ·
We will tell you in the assessment which of your document problems chat solves and which it does not, before anybody signs anything.