An internal policy assistant
Help employees find approved procedures and guidance, with links to the relevant passage and a route to the policy owner for interpretation.
Your company knowledge.
Finally, within reach.
Policies in one folder. Project notes in another. The answer somewhere in between. We build company knowledge assistants using retrieval-augmented generation, or RAG, to help your team find and use the information your business already has.
Discuss a knowledge assistant- Answers linked to sources
- Access designed around your team
- Tested with real questions
An answer is only useful if you can trust where it came from.
A general chatbot does not automatically know your latest procedures, customer commitments or technical documents. Copying files into a chat window can be a useful experiment, but a shared business assistant also needs document ownership, access controls and a way to keep information current.
RAG retrieves relevant passages from an approved collection before a language model drafts an answer. We design the surrounding system: how files enter it, how permissions are enforced, what evidence is shown and when the assistant should say it cannot answer. Citations help people check a response; they do not guarantee that it is correct.
One useful question. The right knowledge behind it.
Technical knowledge search
Search manuals, product documentation and troubleshooting notes. Keep version information visible so an answer about one product does not silently apply to another.
Project and bid research
Find reusable material in approved project records. Show the supporting source and separate published facts from a new draft that still needs review.
Onboarding support
Give new colleagues a starting point for routine questions while respecting the documents their role is entitled to see.
“Which procedure applies to this job?”
The starting point
A colleague searches several folders, messages an experienced team member and finds two conflicting versions of a procedure.
A better way forward
An internal assistant searches the permitted, current collection, links the relevant passage and flags a conflict for the document owner instead of inventing a resolution.
- 1
Prepare the sources
Choose a controlled document set and agree owners, versions and user groups.
- 2
Retrieve the evidence
Find useful passages with access restrictions applied before they reach the model.
- 3
Check the answer
Test correct answers, missing evidence, conflicting documents and prohibited access.
An illustrative engagement. We agree the actual scope, access and success measures with your business.
Something useful.
Yours to put to work.
The engagement has a clear scope and a tangible handover. We agree the deliverables before work starts.
Discuss the scopeA source and permissions plan
Approved collections, update rules and access boundaries, with unsupported sources called out before the build.
A working knowledge assistant
An agreed interface, retrieval pipeline and source references, connected to the systems included in the scope.
An evaluation set
Representative questions with expected evidence, plus tests for unanswered questions, misleading documents and access isolation.
An operating guide
Instructions for adding, correcting and removing content; monitoring quality; handling incidents; and estimating ongoing usage costs.
How we work with you.
Choose a collection
Start with a defined team and a document set with an accountable owner.
Establish a baseline
Record how people answer representative questions today and what good evidence looks like.
Build and evaluate
Compare retrieval and answer quality, including deliberate failure and permissions tests.
Release in stages
Pilot with a small group, collect corrections and agree who maintains the sources.
A few things
you might be wondering.
Questions about your own setup?
Let’s talk it through
What is RAG development?
It is the design and implementation of a system that retrieves relevant information and gives it to a language model as context for an answer. Business RAG development also includes document ingestion, permissions, evaluation, source updates and the user interface.
Is RAG the same as training an AI on our documents?
No. A RAG system typically searches your content at answer time; it does not need to change the model’s learned weights. That makes retrieval a useful option to investigate for changing company information. Fine-tuning addresses different requirements, such as consistent task behaviour, and can sometimes be combined with retrieval.
Can it search SharePoint, Google Drive or PDFs?
Those are possible sources, subject to the access methods, licences, file quality and permissions available. We check each proposed connector and test representative files before promising a particular integration. Scanned documents may need an extraction stage.
Will everyone be able to see every document?
That should depend on your access policy. We define user groups and retrieval restrictions, test them with separate accounts and include deletion and permission-change behaviour in the design. A model prompt alone is not an access-control system.
How do you measure whether the assistant works?
We assess whether it retrieves the right evidence, answers the question accurately, cites a supporting passage and declines questions it cannot substantiate. We also measure response time, review effort and cost using representative tasks.
What affects the cost of a RAG project?
The main variables are the number and condition of sources, permissions complexity, update frequency, interface requirements, evaluation effort and expected usage. We scope a limited collection first and separate implementation from hosting, model and maintenance costs.
A little clarity before the next step.
Make the knowledge you already have easier to use.
Tell us what you’re working on. We’ll help you find a practical next step.
Discuss a knowledge assistant