AI · Generative AI & LLM Applications
An answer is only as good as what it was allowed to read.
Noble builds applications on large language models that an organisation can put in front of its own people: assistants inside the tools they already use, knowledge systems over the documents they are permitted to see, and generation that carries its sources with it. The model is bought from the market and will change; the retrieval, the permissions, the evaluation set and the integration are the work, and they are what stays.
Noble does not train foundation models. The work is applied engineering on models that come from the market — choosing one per engagement, grounding it in the organisation’s own material, and building what surrounds it.
What separates a system from a demonstration
A demonstration answers a question well in front of an audience. A system answers it well for a person who was not in the room, about a document that was added last week, without revealing a document that person is not allowed to see. The gap between those two is the whole of the engineering: retrieval that respects the same permissions as the source, context assembled so the model has what it needs and nothing it should not, and a record of what was retrieved and what was said so that a wrong answer can be traced.
The other separation is measurement. Before Noble shows an assistant to anyone, there is a set of questions with known-good answers, including the ones the assistant is expected to get wrong, and the assistant is scored against it. That set is the definition of done, it is re-run every time the model, the prompt or the content changes, and it is what makes "better" a word with a meaning.
The path from a question to a grounded answer
Two gates. The first is a permission filter applied before retrieval, so the model never sees what the person may not. The second checks the answer against its sources before it is shown, and routes what fails to a person.
- 01QuestionFrom the person, in their tool
- 02Permission filterSame identity as the source
- 03RetrieveGoverned, indexed content
- 04Assemble contextWhat is needed, nothing more
- 05GenerateModel chosen per case
- 06Check against sourcesCited, or routed to a person
What the service covers
Retrieval over an organisation’s own content
Retrieval-augmented generation built as a search system first: content chunked and indexed with its permissions and its provenance, retrieval tuned against the evaluation set, and the retrieved passages shown with the answer so the reader can check. Arabic and English content are indexed together and retrieved across languages when the question is in one and the document in the other.
- Indexing with permissions and provenance attached
- Retrieval tuned against a known-answer set
- Cross-lingual retrieval between Arabic and English
- Sources shown with every answer
Assistants and copilots inside existing tools
An assistant that lives where the work is — inside the service desk, the document system, the ERP screen, the customer portal — rather than in a separate chat window nobody opens. Built on the application’s own identity, scoped to the task it is for, and designed to hand off to a person cleanly when the question is outside it.
- Embedded in the application, on its identity
- Scoped to a task, with a clean handoff beyond it
- Drafting, summarising and answering over live records
- Bilingual by design, composed for Arabic
Knowledge systems on governed sources
The policies, procedures, manuals, contracts and records an organisation has accumulated, made answerable. The work is mostly in the sources: deciding which are authoritative, retiring the copies that contradict them, and keeping the index in step with the document system so an answer is never given from a superseded version.
- Authoritative sources chosen; contradicting copies retired
- Index kept in step with the document system
- Answers that name the version they came from
- Search and question-answering over one index
Prompt, context and retrieval design as engineering
The instructions, the retrieved passages, the conversation state and the tool results that a model sees are its input, and they are designed and versioned like any other input. Noble treats the context window as a budget to be spent deliberately, keeps prompts under source control with the evaluation results that justified each change, and separates what the model must be told from what it must be able to look up.
- Prompts versioned with the evaluation that justified them
- Context budgeted, not filled
- Structured outputs where a system consumes the result
- Model choice revisited as the market moves
Evaluation before demonstration
A set of questions with known-good answers, built with the people who will use the system and including the cases it should refuse. It is scored before the first demonstration, re-run on every change, and reported to the organisation as the honest measure of whether the system is ready — and of whether the latest model is actually better for this task.
- Known-answer sets built with the users, including refusals
- Scored on every change to model, prompt or content
- Groundedness checked against the retrieved sources
- Results reported as they are, to the organisation
The record, and what the model may not do
Every prompt, retrieval and answer is logged against the identity that asked, so a wrong answer can be traced to what the model was shown. Personal data is handled by a written rule, residency is decided before a model is chosen, and the model has no path to act on a system — that is the agents page, and it is a different design.
- Prompts, retrievals and answers logged to the identity
- Residency and personal-data rules set before the model is chosen
- Read-only by design; action is a separate system
Where this connects
- AI Agents & Intelligent AutomationWhen an assistant has to act rather than answer, it becomes an agent, and the permission and stopping rules on that page apply.
- Conversational, Voice & Multilingual AIThe channel — chat, voice, Arabic and English in one exchange — is designed there; the grounding behind it is designed here.
- AI Engineering, MLOps & GovernanceServing, cost, monitoring and the audit trail that keep an assistant honest after launch are that page’s subject.
The services alongside
- Data & AnalyticsThe governed content layer an assistant retrieves from. Nothing here works over sources that disagree with each other.
- SecurityRetrieval is access control before it is ranking; the permissions come from the same identity system as everything else.
- Software & Web EngineeringAn assistant reaches a person through an application, and that application is built and maintained by this service.
Bring twenty questions.
The questions people actually ask — and the answers they should get — are the first evaluation set. Twenty of them, with the documents the answers live in, is enough to know whether there is a system here.