AI assistants grounded in your own documents, products and policies
- Answers with sources from your own content
- A hand-off to a person, by rule
- Evaluated before launch and after every change
- Built on the cloud and model provider you choose
Use cases we build
Customer support deflection
Answers the repeat questions about shipping, returns, pricing rules and how-to from your help content, and passes everything else to your team with the conversation attached.Internal knowledge assistant
Lets staff ask in plain language across policies, manuals, past proposals and wiki pages, with links to the source document so they can check.Product finder
Helps shoppers narrow a catalogue by need ("a waterproof jacket for cycling under $200") using your product data, not invented specifications.Onboarding guide
Walks new customers or new employees through setup steps and answers questions from the onboarding material you already have.How a grounded assistant works
- Retrieval. Your content is split into passages and indexed. Each question retrieves the passages most likely to contain the answer.
- A constrained answer. The model is told to answer only from those passages and to cite them.
- Citations. Every answer links to the page or document it came from, so the reader can check it.
- Refusals. When the passages do not contain the answer, the assistant says so instead of guessing, and it declines topics you rule out (legal advice, pricing promises, other customers).
- Hand-off. Rules decide when a person takes over: an upset customer, a refund above a threshold, two failed answers in a row.
Our buyer's guide to RAG assistants explains the choices and the failure modes in more depth, and how to build a customer-support chatbot walks through the procedure.
Stack options and how we choose
We work with managed and self-hosted components and pick per project:
- Managed retrieval: Amazon Bedrock Knowledge Bases handles ingestion, indexing and retrieval for content in S3, SharePoint, Confluence and similar sources.
- Models: Anthropic Claude, OpenAI models, other models available on Amazon Bedrock, or open-weight models hosted on AWS.
- Vector stores: pgvector on PostgreSQL when you already run Postgres, OpenSearch when you need its search features at scale.
The choice follows four questions: where your data must stay, how many conversations to expect, which cloud you already use, and how much of the system your team wants to run itself. We keep the model behind a thin interface so it can be changed when prices or quality move.
Quality and safety
An evaluation set
Questions with known correct answers and sources, including questions the assistant must refuse and attempts to make it misbehave, run before launch and after every change.Prompt-injection defences
Retrieved content and user messages are treated as untrusted, following the OWASP Top 10 for LLM Applications; instructions found inside them cannot change the assistant's rules.Personal data handling
Personal data stays out of the general index, retention is set on purpose, and the privacy policy says what the assistant does. See privacy for AI features in Canada and India.Logging and review
Conversations are logged with the passages used, so someone can read a weekly sample, fix the content gaps and see where hand-offs happen.What it costs
There is no honest single price, so we quote after scoping. What drives the build cost: how many content sources, whether the assistant needs to look up orders or accounts, how many languages and channels (website, helpdesk, internal chat), and how strict the evaluation needs to be.
Running costs have three parts: model usage, which grows with the number and length of conversations; retrieval and vector-store hosting, where some managed options carry a minimum monthly charge; and the time of the person who keeps the content current. We estimate all three before you commit, and link to the providers' current pricing pages rather than quoting numbers that change.
What we have built
We have built a retrieval-grounded (RAG) assistant, an MCP server that lets AI applications use a business system through defined tools, embeddings-based semantic search, and an AI workflow agent whose actions a person approves. We describe them in general terms here, without client names or numbers; ask us for the details on a call.
The rest of our background is the systems AI has to plug into: we have built websites, online stores and mobile apps since 2005, today mostly on Laravel, Node.js, Magento, Shopify, WooCommerce and AWS, and in October 2026 we moved our own site from WordPress to Astro on AWS.
Further reading from our team:
What we will not do
The same rules apply to every AI project we take on. They are the reason some projects end with a smaller scope, and a few with a recommendation to wait.
- Launch an assistant without telling people they are talking to an AI.
- Put an agent into production that acts without a person who can approve, stop and undo what it does.
- Train or fine-tune a model on your data, or send it to a provider that does, without your written consent.
- Publish AI-generated text on your site or send it to your customers without a person reviewing it first.
Frequently asked questions
What is a RAG assistant?
A retrieval-augmented generation (RAG) assistant looks up relevant passages in a knowledge base you control, such as help articles, product data or policies, and gives them to a language model with instructions to answer only from them and to cite them. It answers from your content instead of from what the model happened to learn.
Can the assistant answer questions about orders or accounts?
Yes, through an authenticated lookup with fixed rules rather than free-text search. Personal data stays out of the general knowledge base, and the assistant only sees the record of the signed-in customer.
How do you stop it making things up?
The model is instructed to answer only from the retrieved passages, to cite them and to say when they do not cover the question. An evaluation set checks this before launch and after every change. No method removes errors completely, which is why every assistant has a hand-off to a person.
Which AI model do you use?
We choose per project from hosted models such as Anthropic Claude and OpenAI models, models on Amazon Bedrock, or open-weight models on AWS, based on where your data must stay, the volume and the cost. We design the assistant so the model can be swapped later.
Will customers know they are talking to an AI?
Yes. We disclose that the assistant is automated and that a person is available. The EU AI Act requires this disclosure from 2 August 2026, and Canada's privacy regulators list openness among their principles for generative AI.
Scope an assistant for your business
Send us the questions your team answers most often and the content they answer from. We will tell you what an assistant could handle.
