Skip to content
Working in-house systemKnowledge and support

Answering business questions from official guidance

Answers questions from an approved set of official guidance and shows the page each answer came from — and declines when the material does not cover the question.

Relevant to: Teams who answer the same policy or procedure questions repeatedly, and need the answer to come with a source.

The problem

Guidance for starting and running a business is spread across many pages. Finding the relevant page can take longer than understanding the answer once it has been found.

A general chatbot can produce a plausible response without showing whether it came from the official material. For tax and employment questions, the reader needs a source and a clear boundary around what the assistant knows.

How the work was done before

The current alternative is searching the guidance yourself: opening several pages, reading past the parts that do not apply, and deciding whether what you found is current and relevant to your situation.

The alternative BrisAI wanted to avoid is a general chatbot that answers confidently from training data, where the reader has no way to tell a sourced answer from an invented one.

What we built

Fourteen GOV.UK pages covering business setup, tax registration, employment and insurance were selected by hand and stored with their page title and source URL. Manual selection keeps the knowledge base narrow enough to review.

The pages are divided into short overlapping sections. Each section is converted into a numerical representation of its meaning — commonly called an embedding — and stored so relevant passages can be found even when the question uses different wording.

When a question arrives, the assistant searches those passages, writes an answer from what it finds and returns the original page title and URL. If the search does not provide enough support, the assistant declines the question.

See the behaviour

Guided workflow demonstration

Follow one representative item from business input to reviewable output. The sequence explains the work; it does not represent processing speed.

Walkthrough ready.

Business guidance question

Sample data
Question
What steps do I need to take before hiring my first employee?
Knowledge boundary
Selected, approved GOV.UK guidance
  1. 01

    Receive question

    Accept the question without adding assumptions.

  2. 02

    Search approved material

    Search only the selected GOV.UK collection.

  3. 03

    Retrieve relevant passages

    Bring the most relevant passages and their page details together.

  4. 04

    Compose the answer

    Write a concise response using the retrieved material.

  5. 05

    Attach checkable sources

    Return page titles and links with the answer.

Read the complete walkthrough as text

Input: Question: What steps do I need to take before hiring my first employee?; Knowledge boundary: Selected, approved GOV.UK guidance.

  1. Receive question. Accept the question without adding assumptions.
  2. Search approved material. Search only the selected GOV.UK collection.
  3. Retrieve relevant passages. Bring the most relevant passages and their page details together.
  4. Compose the answer. Write a concise response using the retrieved material.
  5. Attach checkable sources. Return page titles and links with the answer.

Output: Answer excerpt: Before the employee starts, register as an employer, check their right to work, arrange PAYE, provide required employment information and make sure the appropriate insurance is in place.; Sources: Employing staff for the first time · Check a job applicant’s right to work · PAYE and payroll for employers; Boundary: If the approved pages do not support the question, the assistant declines instead of filling the gap from general model knowledge..

Human control: A person can open every cited page and check the answer against the source before relying on it.

Guided walkthrough using representative sample data. It demonstrates the workflow’s behaviour; it is not an accuracy benchmark or a client result.

What we have measured

No formal result is published yet. This is a working in-house system available to demonstrate, but it has not been put through a declared held-out evaluation.

When that evaluation runs, what will appear here is the question it answered, the dataset it ran against, the frozen workflow and model versions, and the raw counts behind every figure — not a percentage on its own. The evaluation methodology sets out what we hold ourselves to before publishing a number.

How it fits together

The architecture, for readers who want the technical shape after seeing the business behaviour above.

gov.uk pagesChunk 400/50EmbeddingsQdrantQuestionAgentCited answer
Ingestion runs on demand; the query path is live.

Explore the build

Optional technical depth showing the structure of the real workflow. It loads only when this section is approached or requested.

Type
In-house build
Source material
14 GOV.UK pages
Vector store
Qdrant Cloud
Search index
92 text sections
Test status
Working demo

Where this approach could help

The same approach can help teams search policies, procedures, product information or other controlled document collections where answers need a checkable source and a clear limit.