On-premise enterprise AI

Generative AI that stays inside your network.

Multi-model chat with streaming answers, grounded in your own documents through ExA Insight RAG and wired to your systems through ExA Insight MCP. Prompts, retrieval, and model calls all run on hardware you control.

ExA Insight Starter Pack
ExA Insight MCP
Self-hosted LLM
Web Search Proxy

See the packages How it deploys

Public internetUntrusted. Not a data plane.

One way out, and only if you enable it

Your network

Web Search Proxy
Optional, self-hosted egress
Chat UI
Browser, no thick client
Self-hosted LLM
OpenAI-compatible endpoints
ExA Insight RAG
Ingest, embed, retrieve
ExA Insight MCP
Approved tools and APIs
Identity & SSO
MFA, groups, permissions

At a glance

What it is, and who it is for

Audience: Enterprise users, alongside the IT, security, operations, and platform owners accountable for where data lives.
Scale target: Typical deployments run tens to low hundreds of concurrent users per installation. Sizing varies with hardware.
Core value: Multi-model chat on your infrastructure, answers grounded by ExA Insight RAG, tool-augmented workflows through ExA Insight MCP, and an optional self-hosted Web Search Proxy for fresh web context.
Adoption: Pilot first. Expand domains, integrations, and admin scope while least privilege stays enforced.

Capabilities

From models to grounded answers

Chat is the surface. What makes an answer usable is where the context came from and whether you can check it.

Assistants, not models: Each assistant carries its own self-hosted, OpenAI-compatible endpoint, system prompt, and persona. Users pick an assistant, so one choice sets the model, the instructions, and the knowledge it can reach — together.
Streaming chat: Answers arrive token by token and render as markdown. Where the model supports it, its reasoning shows in a block you can collapse.
ExA Insight RAG: Ingestion, chunking, embeddings, and retrieval across approved corpora, all in-network. Datasets are mapped to assistants, so an answer draws only on the corpora that assistant is entitled to.
Grounding in chat: When retrieval returns context, the answer carries numbered citations and a collapsible source list. Downloading a source runs the same access checks that governed the retrieval.
Web Search Proxy: When on-premise knowledge runs thin, route search through your own self-hosted proxy. Supported cloud search providers stay optional, and off until your policy allows them.
ExA Insight MCP: The same assistant can call approved tools and backends over MCP — search, internal APIs, and future connectors — each under explicit permission.
Conversation modes: Persistent threads keep history and titles. Temporary chat leaves nothing behind, for one-off exploration.

On-premise deployment
Docker Compose
Multilingual UI

Security & compliance

The perimeter is the product

Residency is not a setting to switch on later. It is what the architecture is for.

Your network

Privacy by design: Prompts, documents, and model calls stay inside your network. Nothing is sent to an external cloud API.
Zero Trust access: Every session is verified by identity and device. MFA is enrolled, not offered, and login grants no implicit trust afterwards.
Zero public port exposure: No ports open to the internet. Admins reach the system over encrypted, Zero Trust-verified channels from any network.
Least privilege, enforced in the backend: An assistant may only read the datasets and call the tools mapped to it, and a user only reaches assistants their group allows. The check runs server-side on every call — the model never holds a credential, an endpoint, or the choice of which dataset to search.
On-premise data residency: Inference, retrieval, and the Web Search Proxy all run on your hosts. No workflow moves data across your network boundary.
SSO: Integrates with the identity management you already run, including group claims that map to access inside ExA Insight.

Enforcement details

Session cookies are opaque and validated server-side. No tokens in browser storage.
Passwords are hashed with Argon2id; one-time tokens are hashed and expire.
Permissions are granular and written as resource:action.
MFA reset is admin-initiated. There is no self-service path around it.
An administrator cannot delete their own account or drop their own admin role.

Packages

Start with the stack. Add the bridge when you need it.

Starter pack

ExA Insight Starter Pack

Chat
RAG

A ready-to-deploy on-premise AI stack — conversational chat and grounded knowledge retrieval on your infrastructure, with admin controls that respect how you classify data.

Responsive chat UI without a thick client install
Admin-oriented configuration for users, policies, and knowledge bases
Supports multi-factor sign-in for chat and admin access
Local RAG pipeline: ingestion, chunking, embedding, and indexing stay in-network
Per-domain knowledge boundaries that limit cross-unit leakage
Source-aware context for explainable, accountable answers

Add-on

ExA Insight MCP

Tools
Integrations

MCP servers as the controlled bridge between LLMs and the outside world: internal APIs, optional web search via a private proxy, SSO-routing integrations, and future log or line-of-business data sources — without treating the public internet as a trusted data plane.

Standardized tool surface for APIs, search proxy, and audit-friendly hooks
Explicit tool permissions and least-privilege exposure per connector
Token-aware, rate-limit-friendly patterns for external system calls
Room to grow from pilot automation to broader agent programs under governance

AI administration

Nothing is on until an administrator turns it on

An assistant is not a model with things bolted on. It is one configured object: an endpoint, a prompt, a persona, the corpora it may read, the tools it may call, and the groups allowed to use it.

AI configuration

Build an assistant end to end: its own OpenAI-compatible endpoint, model, system prompt, and persona. There is no shared pool of models — the assistant is the unit.
Map the datasets it may read and the tools it may call. That mapping is its entire reach; nothing outside it exists at run time.
Grant groups access to it. A user sees an assistant only when one of their groups is on that list.
Connect a RAG backend once — retrieval endpoint, download endpoint, credentials. The datasets it exposes become assignable to assistants.
Define tools once — web search through your own proxy, or any generic MCP server — then attach each to the assistants allowed to use it. Supported cloud search providers stay optional, and off until your policy allows them.
Tune long-conversation behaviour per assistant: verbatim window, summary budget, and how many recent turns stay untouched.

Access & accounts

Users, roles, and granular resource:action permissions.
Groups that mirror your departments, with delegated group managers.
OAuth2 and SSO, including group-claim mapping from your identity provider.
SMTP or email-API configuration, test-sent before it is saved.
API keys for embedding chat in your own applications, scoped per assistant.

Oversight

Feedback ratings and per-assistant response performance.
ExA Insight RAG dataset and tool utilisation.
Drill-down from a feedback entry to the conversation itself, behind its own separate permission.

Deployment & licensing

Runs on your hosts, in stages you choose

Deployment: On-premise, typically via Docker Compose. Inference, chat, ExA Insight RAG, and the optional Web Search Proxy run on your hosts. Roll out in stages as entitlements allow.
Licensing: Per-feature entitlements: a basic tier covering LLM and ExA Insight RAG, plus add-ons. Signed licence files are hardware-bound and merged at load, so a failed add-on is skipped while the primary stays valid.
After purchase: Grace windows, how operators see licence status, uploading renewed material, and restart expectations are covered in onboarding and in your order rather than itemised here.

From pilot to production

Scope follows your data, not our feature list

ExA Insight can land as a focused pilot or as a broader program. How far it goes depends on your data readiness, your sensitivity classification, how many integrations you want on day one, and how tightly admin access should mirror least privilege.

Minimum hardware for local AI models

Pilot: NVIDIA DGX Spark
Production: NVIDIA RTX 6000 Blackwell

Minimums, not targets. Final sizing follows your model families, context limits, and expected concurrency.

01
Assess
Curate official sources for ExA Insight RAG, classify sensitivity, define domains and roles, set measurable outcomes, and size inference hardware to pilot or production baselines.
02
Build
Install and configure ExA Insight RAG, chat, local inference on approved GPUs, security components, the Web Search Proxy if needed, and MCP tools for approved workflows.
03
Operate
Monitor quality and usage, tune ExA Insight RAG retrieval, extend MCP integrations, and keep audit and access policies current.

Next step

Talk through a pilot

Scoping starts with your data rather than a feature list: which sources are official, how they are classified, and who should be able to reach what. Bring that, and the rest sizes itself.

sales@cdtech.id

Mention your official sources, how sensitive they are, and roughly how many people will use it — then the first reply can already be specific.