Marcus Biel.

AI search for internal documents

Get answers you can verify in your company documents

A customer asks a question. The product details are in a wiki, the instructions are in a PDF, and the service team has its own notes. Which source has the answer?

I build AI assistants that search the documents each person is allowed to read. Staff get an answer with a link to the passage behind it, so they can check it for themselves.

What does your team need to know?

Ask a few colleagues what they looked up this week. Their questions are a good starting point.

  • Customer service. What should we ask the customer for, and where is that rule written down?
  • Sales. How does this model differ from the one the customer already has?
  • Internal procedures. Which steps do we follow, and who signs off?
  • Onboarding. Where can a new hire find the instructions for their first assignment?

I use those questions to choose the sources, design the answers, and decide how to test them.

How retrieval-augmented generation works

With retrieval-augmented generation (RAG), the application looks up relevant passages and gives them to a language model as the basis for its answer.

I prepare and index the documents, check which ones each user can access, and test whether the cited passage actually answers the question.

Your team can ask questions in a standalone assistant or inside the application they already use.

Integrate a knowledge assistant into a Java application

A service question and the documents behind the answer

A service rep asks two questions. The first answer draws on two approved documents. For the second question, the service team needs a different source. Follow each citation to check what it supports.

Sample questions

What a service rep might ask

“What information do we need to process a service request for product line A?”

Record the serial number, the exact product variant and a short description of the fault. Service guide, Section 2.1

For variant A2, also record the error code on the device. If there isn’t one, say so in the ticket. Variant sheet A2, Section 1.3

“How soon will the customer receive a replacement device?”

Neither approved document specifies a replacement deadline. To give the customer a date, the team needs the service agreement that applies to this order.

View the available documents

Fictional documents · Available to customer service

Source passages

Service Guide · Version 3.2

2.1 · Receiving a request

“For every service request concerning product line A, record the serial number, the exact product variant and a short description of the fault.”

Back to the answer about the service request

Variant sheet A2 · Version 1.1

1.3 · Recording the error code

“For variant A2, also record the error code shown on the device. If no code is displayed, record that in the service case.”

Back to the answer about the A2 error code

  1. 1. Understand the question

    If a detail affects the answer, the assistant can ask for it. Here, the exact product variant determines whether the team needs an error code.

  2. 2. Find relevant documents

    The application searches the service documents and product information the user is authorized to read.

    You decide which document versions count as authoritative for each question.

  3. 3. Present the answer with its sources

    Staff see an answer tailored to their question, with the documents and sections used so they can check the original wording.

    Depending on the source, the answer can also show the document version or when it was last updated.

  4. 4. Handle unanswered questions

    I work with your team to decide what the assistant should do when the documents don’t support an answer. The assistant might ask a follow-up question, show a relevant passage or hand the case to the right person.

    I test what the assistant does in those cases, too.

Which sources can your team rely on?

A knowledge system can use product documentation, internal instructions, approved knowledge articles and selected records from business applications.

Three decisions shape a pilot: which documents are authoritative and who maintains them, who may read them, and which real questions the answers need to stand up to. Start with one team and the sources it uses for a specific job.

A few decisions matter early on.

  • Content owners. Who decides which documents the assistant should use?
  • Authoritative versions. How do we distinguish current, superseded and archived content?
  • Updates. How will the assistant pick up new or revised documents?
  • Removal. What happens when a document is deleted or withdrawn?
  • Source details. Can someone see where a passage came from, which version it belongs to, and when it was updated?

That gives the people who maintain the documents a way to keep the assistant current.

Permissions for each user group

An internal assistant must respect what each person is allowed to read.

The application checks access before it retrieves a passage or includes it in an answer. Each group can search only the sources it has permission to read.

For each integration I check whether permissions can be inherited from the wiki, document management system or other source. Role changes and revoked access are part of the update process.

More about APIs and system integration

Test answers against real business questions

I check whether the assistant finds the information it needs and whether that information supports its answer.

I collect questions from your team, write down what a good answer would contain, and test the assistant against them.

  • Relevant passages. Does the search find the passages that answer this question?
  • Supported claims. Do the cited sources actually support the answer?
  • Completeness. Is anything missing that someone would need to act on the answer?
  • Gaps in the sources. Does the assistant say when the documents don’t settle the question?
  • Conflicting versions. How does it handle conflicting or outdated information?
  • Permissions. Can a user see only the documents and answers they’re entitled to?

I check the answer and its citations together. A link is useful only if the passage supports what the assistant says.

You can run the same questions again after changing the documents, the search, or the model.

Microsoft Learn on evaluating retrieval and answer quality in RAG

Start with one team and one set of documents

A first release works best when it’s clear who will use it, what they’ll ask, and which documents they trust.

For example, the service team could start with the approved documents for one product group. Together we choose the questions the assistant should answer and how to evaluate its responses.

The first release includes

  • An assistant your team can use
  • Connections to the sources you choose
  • Answers with links to the original passages
  • Access checks for each user group
  • Test results for representative questions
  • Notes on running the assistant, keeping it up to date, and its known limits

Once people have tried it, you can decide which documents to add or which team to bring in next.

Plan a pilot with your documents

Planning operations and ongoing costs

The amount of work depends on the source systems, document formats, access rules, and the standard the answers need to meet.

To estimate operating costs, I consider the volume of content, how often it changes, expected usage and the services involved.

I can measure response times, answer quality, and processing costs on real tasks in the first release. I’ll also help you identify who will keep the content up to date and who will provide technical support.

Before I start building, I check your rules for data storage, transfers, and the services allowed to process it.

Plan architecture and operations together

Who leads the work?

I’m Marcus Biel, a software architect and developer. I’ve been building software with Java since 2001.

I’ll work with the people who answer these questions today, design the assistant, and help write the code. I can join your developers or bring in specialists when needed.

Together we’ll decide where staff should be able to ask a question and how they’ll get back to the original document.

More about working with Marcus
Marcus Biel
Marcus BielSoftware developer and architect

What does your team keep having to look up?

Send me a couple of typical questions, who asks them, and where people look for the answers today.

We can use those examples to pick a useful starting point for a pilot.

Prefer to talk first? Call +49 171 1151013 or message me on LinkedIn.

Talk about an AI search assistant