GenAI assistants & copilots

Assistants grounded in your own content, with the evaluation to show whether they are actually right.

Discuss your project Learn this field

What is it?

A demo assistant takes an afternoon. A production one needs grounding in your real documents, handling for the questions it cannot answer, an audit trail, and a way to measure accuracy that is not somebody's impression. We build the second kind.

For a technical reader

The usual architecture is retrieval-augmented generation: documents are chunked, embedded and stored in a vector index; a query is embedded, nearest neighbours are retrieved, and the model answers over that context with citations. Most quality problems are retrieval problems rather than model problems — chunk boundaries that split a table, embeddings that do not capture domain vocabulary, or no reranking step. Tool calling extends this from answering to acting, at which point authorisation stops being advisory and has to be enforced server-side on every call.

Our services

Assistant design and buildAn assistant grounded in your documents and systems, deployed inside your environment with your authentication.
Evaluation harnessA scored question set and automated scoring, so accuracy is a number that moves rather than an opinion.
Retrieval engineeringChunking, embedding, hybrid search and reranking tuned against your actual corpus.
Governance and guardrailsTopic boundaries, PII redaction, refusal behaviour and full prompt/response audit logging.

What’s covered

Retrieval over your content
Chunking, embedding and vector search across your documents, wikis and systems
Tool and function calling
The assistant reads from and acts on your real systems, under explicit permissions
Evaluation harness
A scored question set run on every change, so accuracy is measured rather than assumed
Guardrails and refusal
Topic boundaries, confidence thresholds and routing to a human when it does not know
PII handling and audit
Redaction before anything leaves your boundary, and a log of every prompt and response

Business problems we solve

The pilot demoed well and never shippedDemos skip evaluation, authorisation and audit. We build those first, so there is a defensible path to production.
It answers confidently and wronglyWithout grounding and a refusal path, a model will fill gaps. Citations and confidence thresholds make answers checkable.
Staff cannot find what they needPolicy and procedure spread across wikis, PDFs and shared drives. Retrieval over all of it beats reorganising the intranet again.
Legal will not approve itUsually because nobody can say what data leaves the boundary or what is logged. Those controls are part of the build, not an afterthought.

Our delivery approach

  1. 01
    Pick a question set firstWe agree the questions it must answer correctly before building. That set becomes the evaluation harness and the definition of done.
  2. 02
    Ground it in your contentRetrieval over your documents and systems, with citations, so answers can be checked rather than trusted.
  3. 03
    Design the refusal pathKnowing when to say "I don't know" and route to a human matters more than breadth. A confident wrong answer is the expensive failure.
  4. 04
    Evaluate continuouslyAccuracy is tracked against the question set on every change, so regressions surface before users find them.

Solutions & project types

Examples of the work we take on in this area. These describe project shapes, not named client engagements.

Internal policy assistantAnswering HR, finance and compliance questions from your own policy documents, with citations.
Customer support copilotDrafting agent responses from product documentation and past resolved tickets, with the agent sending.
Bid and proposal supportRetrieving reusable content and prior approved answers when responding to tenders.
Engineering knowledge assistantAnswering questions over runbooks, architecture decisions and incident history.

Technologies & platforms

Models

  • Anthropic Claude
  • OpenAI GPT
  • Azure OpenAI Service
  • Oracle Generative AI
  • AWS Bedrock

Retrieval

  • pgvector
  • Pinecone
  • Elasticsearch
  • OpenSearch
  • Azure AI Search

Orchestration

  • LangChain
  • LlamaIndex
  • Semantic Kernel
  • Direct SDK

Evaluation

  • Ragas
  • promptfoo
  • Custom scored question sets

Operations

  • LangSmith
  • OpenTelemetry
  • Prompt and response audit logging
Learn this field

Where to start with genai assistants & copilots.

Secondary to the work itself, but genuinely useful if you are building the skill. Every link goes to official documentation or an official channel.

Beginner

  • How a language model generates text, token by token
  • Prompting: instructions, examples, output format
  • Embeddings and what vector similarity means
  • Build a small retrieval question-answer app
  • Understand context windows and their limits

Intermediate

  • Chunking strategies and why boundaries matter
  • Hybrid search: keyword plus vector, and reranking
  • Tool and function calling with real side effects
  • Building an evaluation set and scoring runs
  • Cost and latency: caching, batching, model routing

Advanced

  • Retrieval quality diagnosis: is it retrieval or generation
  • Agentic loops and bounding what they may do
  • Fine-tuning versus retrieval, and when each is justified
  • Red-teaming, prompt injection and data exfiltration
  • Production monitoring and regression on model updates

Certifications

  • Microsoft Certified: Azure AI Engineer Associate
  • AWS Certified Machine Learning Engineer – Associate
  • Google Cloud Professional Machine Learning Engineer
  • Oracle Cloud Infrastructure Generative AI Professional

Practice projects

  • Build retrieval question-answering over a folder of your own PDFs
  • Write 50 question-and-answer pairs and score your assistant against them
  • Add citations so every answer links to its source passage
  • Add a refusal path when retrieval confidence is below a threshold
  • Try to prompt-inject your own assistant, then close the hole
FAQs

Questions we get asked.

Will our data be used to train a model?

Not under the enterprise terms we deploy on. Which provider, which region and what is retained are decided explicitly during the build and written down.

Can it run entirely inside our environment?

Retrieval and orchestration, yes. The model itself can be a hosted endpoint in your cloud tenancy or, with open-weight models, self-hosted. There is a real cost and quality trade-off and we will set it out.

How accurate will it be?

Nobody can answer that before seeing your corpus. That is why the first deliverable is an evaluation set: it turns the question into a measurement you can track.

Do we need to reorganise our documents first?

Usually not. Retrieval works over messy corpora. Genuinely contradictory documents are a real problem, but that is a content issue the assistant will surface rather than cause.

Is this a chatbot?

A chat interface is one option. Often the better placement is inside an existing workflow — drafting a response, pre-filling a form, summarising a case — where it saves time without anyone having to visit a separate tool.

AI & Data

Related services.

Forecasting & machine learningModels that stay in production and keep being retrained, not a notebook that was accurate once.Data platform & warehousingOne place where the numbers agree, with lineage back to the source.Document & contract intelligenceExtracting structured, checkable data from contracts, invoices and forms at volume.Data governance & guardrailsKnowing what data you hold, who can reach it, and what the AI systems are allowed to do with it.
Start the conversation

Talk to our experts about genai assistants & copilots.

Tell us where you are. A specialist will read it and reply within one business day.

Discuss your project