Modules
Modules, not sprawl
Split knowledge by topic, tenant, or time. Each module is self-contained, so adding to one never slows the rest, and a query is routed only to the modules that matter.
How modules work Paddock · The Knowledge Plane for AI
The Knowledge Plane for AI
Paddock is the Knowledge Plane for AI. Your knowledge lives in Knowledge Modules, each one permissioned on its own. You bind a reasoning model on top, and it sees only what its modules allow. When a better model ships, bind that one instead. Nothing in the plane changes.
BOUNDED AI ARCHITECTURE
Three layers, each bound to something you control
Paddock is the Knowledge Plane of the Bounded AI Architecture (BAA). Three layers, each with one job: the knowledge never moves, the model can, and the log covers both.
- 01 · The reasoning layer
Ram
A frontier model compressed to the size you name, reasoning kept and certified, bound to the plane. Swap it as models improve.
How Ram works - 02 · The Knowledge Plane
Paddock
Your knowledge bound to Knowledge Modules: one per topic, period, person, permission or tenant. Permission, freeze, hand back or delete each one on its own.
How modules work - 03 · The governance layer
Watchman
Every module permission, every bind and every model's provenance bound to one log. Five certificates before anything goes live.
What Watchman governs
Because the model holds no facts of its own, swapping it changes nothing in the plane. Your modules, their permissions, and every citation stay exactly where they were.
A wrong answer from a document assistant is worse than no answer, because someone acts on it. Paddock is built for the questions where the number has to be right.
Answers you can stand behind
What Paddock Does
Every answer cites its page
Every answer cites its page
Each answer links to the page it came from. Your engineer clicks through and reads the source in one step, instead of trusting a paragraph and hoping. Citations open the original document at the right place, figures included.
Table-precise, not table-adjacent
Table-precise, not table-adjacent
Ask for a torque value, a weight limit, a fault code, a part number. Paddock returns the exact figure from the table with its page, not a paraphrase that rounds it or reads the wrong row. This is where generic document search quietly fails, and where technical teams get burned.
Grounded, or it tells you
Grounded, or it tells you
Paddock answers from your documents and cites them. When the answer is not in your material, it says so rather than inventing one. For a maintenance manual or a policy set, an honest "not found" is the safe answer, and the one your auditors want.
Diagrams answer too
Diagrams answer too
Schematics, wiring diagrams, and charts are indexed alongside the text. A question about the forward power schematic returns the figure, as a thumbnail you can open, not a sentence describing where it might be.
Knowledge Modules
Knowledge Modules, permissioned one by one
A Knowledge Module is a complete index of its own. Draw one around a topic, a period, a person, a permission or a tenant. Permission it, freeze it, hand it back or delete it. The model bound on top sees only what the module allows.
By Topic
One module per domain
Keep product lines, sites, or subjects in their own modules. A question about one never drags in the noise of another, so answers stay sharp as your library grows.
By Time
Ask as of a date
"What did the maintenance manual require on the day of the incident?" Paddock answers from the revision that was current then, not the one you hold today. Point-in-time recall for aviation effectivity, insurance claims, and legal discovery, where the date is the whole question.
By Person
One module per individual
A patient, an employee, a policyholder. Their records live in a module of their own, reachable only by the roles you grant, and removable in one operation the day the obligation ends.
By Permission
Walls a query cannot cross
Access is granted module by module. A team outside the grant has no index to search, so there is nothing to leak. Ethical walls, need-to-know, and clean teams become structure, not policy.
By Tenant
One deployment, strict walls
Serve many customers or business units from a single install, with hard separation between them. Each module holds its own knowledge and no one else's, so nobody's query reaches another's documents.
Under the Hood
Every module is its own index
A module is not a filter over one big store. It is a complete, self-contained index you can permission, freeze, hand back, or delete, and show the log for all four. How modules work.
A new module does not mean a new system. Add a topic, a person, a time window, or a tenant, and Paddock keeps the rest of your knowledge exactly where it was.
The Reasoning Layer
Bind the model you choose. Swap it later
Paddock holds the knowledge. The model bound on top only reads it. Bind Ram, our frontier models compressed to the byte with reasoning kept and certified, or a cloud model, or your own. When a better model ships, bind that one. Your modules, permissions and citations do not move.
Start on a cloud model, or keep the whole thing private. Same plane, your decision.
Run it on-premise. When the data cannot leave, Paddock runs a Ram model on hardware you control, air-gapped, with nothing going out. A Mac appliance that runs everything on one machine, in a hangar, a branch, a ship, a bank. This is the shape we specialise in.
Run it in your cloud. Containers your platform team drops into its own stack as infrastructure as code. Your cloud account, your region, your keys. Your library and indexes live on volumes you own, and can sit in whatever residency the rules require.
Or let us run it. On our managed SaaS you get the same plane with the same modules, one isolated module set per tenant, and you are answering questions in days. Because modules are portable, starting on SaaS never locks you in: move on-premise later and take your modules with you.
| On-premise | Mac appliance or your servers · air-gapped · 0 outbound |
|---|---|
| Your cloud | Containers, infrastructure as code, your VPC and keys |
| SaaS | Managed by us, one isolated module set per tenant |
| The reasoning layer | Ram on your hardware, a cloud endpoint, or your own model. Swappable |
| Portability | Modules move with you, SaaS to cloud to metal |
SaaS gets you started. Private keeps the data in. We build for the second one, and we will tell you which your case needs.
Ingest
Feed it what you already have
Point Paddock at the documents you keep today. It reads them where they are, keeps web sources in sync, and gives every source its own settings.
| Source | What it handles |
|---|---|
| Manuals and reports, including dense technical tables | |
| Websites | Support portals and docs sites, crawled and re-synced on a schedule |
| HTML / Markdown / text | Knowledge bases and internal notes |
| CSV / JSON | Structured records and exports |
| You get back | |
|---|---|
| Exact value | The figure from the table, with its page |
| Grounded prose | An answer built from your passages, cited |
| Figure | The diagram itself, as a clickable thumbnail |
| Honest miss | "Not in these documents," when it isn't |
Beyond Answers
Turn your knowledge into decisions
Retrieval is where it starts. Paddock lets your people interrogate the whole corpus, weigh what different sources say, and surface the evidence behind a call, without a document ever leaving your control.
Ask across everything at once
Put one question to every module you are allowed to see. Paddock gathers the relevant passages from every source that bears on it and lays out what they actually say, so an analyst starts from the evidence instead of hunting for it.Evidence, not opinion
Every claim carries the citation it came from. A recommendation traces back to the exact page that justified it, so the people who sign off can see the ground it stands on.The reasoning stays home
The analysis runs where your data lives. Your knowledge is never shipped to someone else's model to be reasoned over, unless you decide it should be. The intelligence comes to your data, not the other way round.
The institutional memory your organisation spent decades building finally goes to work for you, on your terms, inside your walls. It is how we run our own lab: read how our research programme lives on Paddock.
Data Sovereignty
One module per person, when you need it
Regulations like GDPR do not ask you to try. They ask you to prove. Paddock gives every individual, team, or customer their own Knowledge Module, so personal data is isolated by construction rather than by policy.
- Isolation
Separated by construction
Each person's data sits in its own module. A query inside one can never reach another, so there is no shared index quietly mixing records that should never meet. - Erasure
Right to be forgotten, cleanly
A patient withdraws consent, an employee departs, a retention clock runs out. Delete that individual's module and their data is gone in one operation, with nothing left bleeding into a shared store. The erasure is complete because the separation was real to begin with. - Residency
Point at where it lives
Keep each module on the hardware, in the region, or under the tenant the rules require. Data residency stops being a promise and becomes an address you can show an auditor.
Privacy that is designed in reads very differently to a regulator than privacy that is bolted on. The full lifecycle story, permission it, freeze it, hand it back, delete it, is on the Knowledge Modules page.
Scale
Built to hold your whole organisation
Paddock grows module by module. Your knowledge splits into independent Knowledge Modules, and a question only touches the ones that hold the answer, so recall stays fast and sharp whether the library is a thousand pages or ten million.
Patent-Pending
Patent-pending binary search
Our patent-pending binary search finds the right passages using up to 60x less storage than a traditional RAG system. The same knowledge fits on a fraction of the hardware, so a corpus that used to demand a cluster runs on a box you already own.
How Paddock search works 
Fast
Fast at any size
Retrieval time holds steady as the library grows. Paddock searches modules in parallel and skips the ones a question does not need, so the last document lands as quickly as the first.
See the measurements Less storage, faster answers, and a footprint small enough to keep the whole thing behind your own firewall. Modules are also how Paddock is governed: one index per patient, client, matter, or year, each with its own permissions and lifecycle.
Architecture
Built over years, not bought off the shelf
Paddock is one engine, tuned end to end. The models, the search, and the index were built together over years of our own research, with Watchman keeping the log. That is why the answers land where generic tools guess.
Merino Models
Our own embedding and rerankers
Merino is our family of embedding and reranker models, built and tuned for the documents enterprises actually keep: dense tables, technical manuals, contracts. Generic models were trained on the open web. Ours were shaped for this.
Binary Search
A search we designed
Our patent-pending binary search is why Paddock stays fast and light as the library grows. How it works stays ours. What it gives you is exact answers on a fraction of the hardware.
Performance
Measured, not asserted
On our own benchmarks Paddock puts the right source in its top results more than 95% of the time, and says so plainly when the answer is not there. Every claim here traces to a number we can show you.
This is not a wrapper around someone else's API. It is years of research, our models, our search, and our index, tuned together. You will not assemble the same result from parts on the shelf.
Stand up your Knowledge Plane
Bring us a manual set, a policy library, or a support site. We will stand up a governed, cited answer engine on hardware you control, with the reasoning model of your choice bound on top.