AI solution design · Knowledge and RAG
Policy and procedure assistant
Staff ask a question in plain English and get an answer from the current policy, with the document and section alongside.
Talk about something like thisThe situation
A company has several hundred policies, procedures and forms spread across SharePoint and shared drives, some of them out of date. Managers and new starters ask HR and IT the same questions every day, and the answer depends on which version of which document you find.
How it works
Step by step
- 1
Collect
Connect to SharePoint and the file shares, pull the text out of Word, PDF and web pages, and record who owns each file, when it changed and who is allowed to read it.
- 2
Prepare
Split documents into passages that make sense on their own, keeping headings and version numbers. Superseded versions are left out, so the assistant cannot quote a policy that no longer applies.
- 3
Search
A question is matched by keyword and by meaning, and results are trimmed to what the person asking is permitted to see.
- 4
Answer
The model is given the best passages and told to answer only from them. It cites the document and section, and says so when it cannot find an answer.
- 5
Check
A test set of real questions, agreed with HR and IT, is re-run after every change so that quality is measured and not assumed.
What it is built from
- Azure AI Search
- Keyword and vector search, with permissions applied to results
- Azure OpenAI
- Writes the answer from the retrieved passages
- SharePoint connector
- Keeps the index in step with the documents
- .NET API and Angular front end
- The chat page, or a Teams app, and the admin screens
- Microsoft Entra ID
- Sign-in and document-level access
Safeguards
- Answers only from retrieved documents, never from the model's general knowledge
- Every answer cites its source, and "I can't find that" is an acceptable reply
- People only see answers drawn from documents they could open themselves
- Questions that go unanswered are reviewed, so gaps in the documents get fixed
What you end up with
- A chat page or Teams app for staff
- An admin view of unanswered and badly rated questions
- A map of which documents are used, and which are out of date
- The test set, so you can keep checking quality afterwards
Where to start
Begin with one department and one document set, such as HR. A first working version can usually be put in front of real users within a few weeks.
More in this group
Support knowledge copilot
A panel inside the ticket screen that suggests an answer from manuals, release notes and past tickets, for the agent to check and send.
Read the design →Contract search and Q&A
Find renewal dates, notice periods and liability caps across hundreds of contracts, and ask questions with the clause shown in the answer.
Read the design →