Skip to content

Service · Private LLM

Private LLMs and Knowledge Assistants on Your Infrastructure

Give your teams the convenience of a ChatGPT-style assistant that answers from your own policies, manuals and records, cites its sources, respects who is allowed to see what, and runs on infrastructure you control.

The problem

Your teams already use AI assistants. The question is which ones.

People want fast answers: what does our leave policy say, which clause covers this, what is the process for a high-value refund. When the internal answer is buried in a 200-page PDF, they open a public AI tool and paste in whatever they have. That is how confidential policies, client details and internal numbers end up on infrastructure your organisation never approved.

Banning AI tools rarely works. It pushes the behaviour out of sight. The better answer is to give people an assistant that is just as easy to use, but that is grounded in your approved content and runs under your control.

Capabilities

What we build

A private knowledge assistant is more than a model on a server. It is the model plus the retrieval, permissions, logging and evaluation that make it trustworthy.
  • Retrieval over your documents

    The assistant searches your approved documents for each question and answers from what it finds (retrieval-augmented generation, or RAG).

  • Source citations in every answer

    Each answer links to the passages it used, so people can check it in one click.

  • Role-based access

    People only get answers from documents they are already allowed to see. HR files stay with HR.

  • Usage and audit logs

    Questions, sources and answers are logged inside your environment for review and improvement.

  • Model selection and evaluation

    We test candidate models on your real questions and pick the one that balances quality, language support and hardware cost.

  • Hardware and hosting sizing

    We size the servers or cloud instances you need, so you are not paying for capacity you will not use.

How it works

How it works

Every question is answered from your own content, inside your own environment.
Inside your environment
  1. 01

    Question

    An employee asks a question in plain language, in a chat window or inside a tool they already use.

  2. 02

    Access check

    The system confirms who is asking and which documents they are allowed to see.

  3. 03

    Retrieve

    It finds the most relevant passages from your indexed documents.

  4. 04

    Private model answers

    A model running on your infrastructure writes an answer using only those passages.

  5. 05

    Answer with sources

    The reply shows its citations, and the exchange is logged for review.

Public AI cloud

not in the data path

Retrieval-augmented generation (RAG) inside your environment

RAG in plain English

A language model on its own only knows what it learned during training, which does not include your policies or your clients. Retrieval-augmented generation fixes that without retraining anything. When someone asks a question, the system first looks up the most relevant passages in your documents, then hands those passages to the model with an instruction: answer using this material, and say where you found it.

Because your documents are looked up at question time rather than baked into the model, you can add, update or remove content at any time, and access rules are enforced on every question.

How we keep answers grounded

Language models can produce confident answers that are wrong. We reduce that risk in layers: answers must come from retrieved passages, every answer shows its sources, the assistant is instructed to say when it cannot find an answer, and we test it against a set of real questions with known answers before launch. For high-stakes uses, answers can be routed to a person for review.

Choosing a model

We work with open-weight models that can run inside your environment, selected for your languages, accuracy needs and hardware budget. The model is a replaceable component: as better models are released, you can switch without rebuilding the system.

Data residency

Where your data lives

The model, the document index, the conversation logs and the access rules all run in the environment you choose: on-premise, in your private cloud or on an India-region cloud. Your documents and questions are never used to train our models or anyone else's.

On-premise

Runs on servers in your own building or data centre. Maximum control; you provide or buy the hardware.

Best for: highly sensitive data, air-gapped needs

Private cloud

Runs in your own cloud account or a dedicated tenancy that only your organisation uses.

Best for: teams already on a cloud platform

India-region cloud

Runs on cloud infrastructure located in India, chosen by you, with data stored and processed in-region.

Best for: fast start with data residency in India

Read more about our commitments on the Trust, Security and Data Residency page.

Engagement path

From risk check to running system

  1. 1

    Risk Check

    Free · 30 minutes

    We map where AI is used today and where your data goes.

  2. 2

    Audit

    Fixed scope

    A structured assessment of your AI use, data flows and gaps, with a prioritised plan.

  3. 3

    Pilot

    One use case

    One high-value use case, deployed in your environment, with clear acceptance criteria.

  4. 4

    Deploy & govern

    Ongoing

    Production rollout with monitoring, access controls, human review and ongoing support.

FAQ

Questions about Private LLM & Knowledge Assistants

How is this different from ChatGPT Enterprise?

ChatGPT Enterprise is a hosted service run by OpenAI. A private assistant runs on infrastructure you choose, uses models you control, and keeps the documents, index and logs inside your environment. Which one fits depends on your data, your regulators and your risk appetite; we compare both honestly during the audit.

Which models do you use?

We use open-weight models that can run inside your environment, chosen after testing on your real questions for quality, language support and hardware cost.

What hardware do we need?

It depends on the number of users, the model size and response-time expectations. We size it for you, and an India-region cloud is often the fastest way to start before deciding on on-premise hardware.

Can it answer in Hindi or Marathi?

Yes, multilingual models can understand and answer in Indian languages. Quality varies by language and task, so we test with your questions before launch.

How do you stop it from making things up?

Answers must come from retrieved passages and show their sources, the assistant is instructed to say when it does not know, and we evaluate it against known questions before launch. Human review can be added for high-stakes answers.

Will people see documents they should not?

No. Access rules are checked on every question, so the assistant only retrieves documents the person is already permitted to read.

Find out where your data is going, in 30 minutes

Book a free AI Data Risk Check. We'll map where AI is used in your organisation today, flag the biggest risks, and send you a one-page summary. No obligation.