Question 1

What is a RAG chatbot and how is it different from ChatGPT?

Accepted Answer

RAG stands for retrieval-augmented generation. Before answering, the assistant searches your own documents, passes the most relevant passages to the language model and asks it to answer from those passages only. A general chatbot answers mainly from what it learned in training. A RAG assistant answers from your current policies and manuals, can show the source, and can say when the documents do not cover a question.

Question 2

How do you stop the chatbot from making up answers?

Accepted Answer

We cannot make a language model perfect, but we reduce the risk in layers. Retrieval is tuned so the right passages are found, the prompt restricts answers to that text, every answer cites its sources, and the assistant is told to decline when nothing relevant is found. We then test against a set of your real questions and review failures before launch and after each change.

Question 3

Is our data used to train OpenAI's models?

Accepted Answer

Under OpenAI's and Microsoft's published terms for their APIs, prompts and outputs sent through the OpenAI API or Azure OpenAI are not used to train their models by default, though both may keep data for a limited period for abuse monitoring. Azure OpenAI also lets you choose the deployment region. If documents must not leave your servers at all, we can run an open-weight model locally instead. We explain the trade-offs in writing so your team can decide.

Question 4

Can the chatbot run on our own servers without a cloud AI provider?

Accepted Answer

Yes. Open-weight models can be served on your own hardware behind an OpenAI-compatible API, using LM Studio for trials or a server such as vLLM for production, with Qdrant for search, so nothing leaves your network. Local models need a suitable GPU or a capable workstation for good response times, and smaller models are weaker at long or complex answers. We usually test a local model against your question set before recommending it.

Question 5

Which documents and sources can it answer from?

Accepted Answer

PDF, Word, PowerPoint, plain text, HTML pages and wiki or CMS content are the common ones. Scanned files need OCR, and spreadsheets or database records often need a different approach from text retrieval. We can sync from shared drives, SharePoint, a CMS or a database on a schedule. A sample of your documents tells us quickly what extra work is needed.

Question 6

How much does AI chatbot development cost?

Accepted Answer

The build is quoted at a fixed price after we have seen sample documents and your questions. The main drivers are the number and quality of sources, how often they change, permission rules, integrations such as single sign-on, and how much testing a wrong answer justifies. Running costs are separate: model usage per token, hosting and search. We estimate both before you commit.

Question 7

How long does it take to build a knowledge assistant?

Accepted Answer

A pilot over one clean document set is a focused piece of work, usually in the two to six week range of our other focused integrations. A production assistant with several sources, sync, single sign-on and permission rules takes longer. We deliver in two-week increments on staging, so you are testing real answers from early in the project.

Question 8

Can you add a chatbot to our existing website, portal or app?

Accepted Answer

Yes. We expose the assistant as an API and add a chat component to your existing Angular, React or Next.js application, or embed a standalone widget. It can use your current login so answers respect each user's permissions. If you already have an internal tool or intranet, the assistant usually fits there better than in a separate app.

Software, cloud and AI systems, built to last.

Nine disciplines, one engineering team.

Systems in production.

We also build products of our own.

A written architecture before any code is written.

The technologies we use, organised by layer.

IT systems that organisations depend on every day.

Our values

The sectors we work in.

Services built around your operations.

Looking for something specific?

{{ svc.title }}

Specific services

Related projects

{{ land.title }}

{{ land.problemsTitle }}

{{ land.includesTitle }}

{{ land.plansTitle }}

{{ land.stepsTitle }}

{{ land.whyTitle }}

Related work

Questions we get asked.

Industries we build for.

Our work, from 2020 to today.

{{ proj.title }}

Platforms we run. Libraries we publish.

Five stages, from first conversation to production.

Engagement models

Notes from the engineering team.

{{ post.title }}

{{ b.x }}

Common questions.

Tell us what you are building.

{{ legalTitle }}

{{ s.h }}

{{ ctaLead }} Let's talk.