Solutions · Chatbots · Custom project Knowledge chatbots for your documents and workflows.
Together with you, we develop an application that finds the right information and supports the steps you agree on.
The answer is in your documents. But where?
Policies, manuals, contracts: the knowledge is there. But before anyone finds the right passage, there are searches, follow-up questions and emails to the specialist team.
The challenge: Knowledge spread across many places
Policies sit on the intranet, manuals on a shared drive, contracts in the archive. A keyword search for “leftover vacation” misses a document that says “unused vacation”.
With deepsight: One question across all sources
The knowledge chatbot searches all the documents you approve – for the meaning of the question, not just the keyword – and answers in full sentences.
How the search worksThe challenge: The same questions, again and again
HR, IT service or sales support answer questions every day whose answers have long been written down in a policy. That leaves less time for the tricky cases.
With deepsight: Ask the chatbot first
Employees ask the knowledge chatbot first instead of emailing the specialist team. It answers from the same documents the team would point them to.
Typical questionsThe challenge: Answers without evidence
A general AI chat doesn’t know your current policies. Its answers sound plausible but can’t be checked against your own documents.
With deepsight: Answers with evidence
The knowledge chatbot relies on your documents and names the document and passage. That way, anyone can look up what an answer is based on.
See the exampleThe challenge: Confidential documents
Pay rules, contracts or internal instructions aren’t meant for everyone. And IT and data protection rightly ask where the documents are processed.
With deepsight: Access and operation on your terms
We agree with you who can reach which documents through the chatbot, and set up the infrastructure with your IT – including local language models in your environment if you wish.
Collaboration and technology
An answer you can verify
Three everyday questions, three documents. Choose a question.
Works agreement on vacation · § 4 · as of 01/2026
(1) Unused vacation of up to five working days may be carried over to the following calendar year.
(2) Carried-over vacation must be taken by March 31 of the following year.
(3) Different rules for parental leave or extended illness are set out in § 7.
Select a sentence in the answer to see the source passage.
Illustrative documents. Check important details at the passage cited.
Not to be confused with the deepsight cloud assistant: it answers questions about your analysis, while a knowledge chatbot answers questions about your own documents.
Finding and acting stay separate
The knowledge chatbot searches the documents you specify and points to the passage. If the documents contain no suitable information, it should say so instead of guessing; how reliably that works is something we test together using your questions.
Suggestion: create a service desk ticket
Illustrative example. Only steps you define in advance are carried out, and only after confirmation if you wish.
How a project works
- Clarify the use case
Which questions, which documents, which steps – and who will use the knowledge chatbot?
- Prototype
A first version using a selection of your documents.
- Test with the specialist team
Check and refine answers and sources using real questions.
- Operation
We agree on access, document updates and further development together.
Questions your documents can answer
Examples from HR, IT, sales, service, purchasing, quality and market research.
- Which documents do I need to apply for parental leave?
- How do I set up VPN access on a new laptop?
- Which services are included in the Basic plan, and which only in the Plus plan?
- What does error code E42 mean according to the maintenance manual, and what should I do?
- What liability cap applies to late delivery under the framework agreement?
- Which inspection step does the work instruction require before release?
- What did our past study reports find on price perception, and where does it say so?
Good to know: Anything that requires weighing up an individual case stays with your specialist team.
What the knowledge chatbot works with
The documents your team already works with today – usually PDF and Word files or pages from your intranet and wiki.
Good to know: Scanned documents need text recognition. How well it works depends on the quality of the scan.
A knowledge chatbot is only as good as the documents it can access.
How an answer is created
Our knowledge chatbots use retrieval-augmented generation (RAG): search first, then write.
- Prepare the documents
The specified documents are split into sections and indexed for search. Title and section are kept so the answer can cite the source passage.
- Find relevant passages
For each question, the application finds the best-matching sections and ranks them by relevance. They form the basis of the answer.
- Write the answer
A language model receives the question together with these sections and writes the answer from them, citing the document and passage.
RAG in the glossary - Apply rules
Depending on the project, agreed rules apply before the answer is shown – such as limiting it to certain topics or referring people to a contact person.
What you have at the end
Start with one area and real questions – not a chatbot for everything.
Start where the same questions come up again and again, then expand.
What matters with knowledge chatbots
Split documents sensibly
How documents are split determines whether tables, footnotes and exceptions stay with the paragraph they belong to during search. We adapt this to your document types.
Only valid versions
If there are several versions of a policy, the search may find the outdated one. So only approved versions go into the collection, and the source shows the version date if the document has one.
Data protection for questions, too
Questions can contain personal data. Whether and for how long conversations are stored is something we decide with the people responsible for data protection at your organization. If we process data on your behalf, we sign a data processing agreement under Art. 28 GDPR.
Questions about knowledge chatbots
No, there is no guarantee. Even with the right documents, a language model can overlook an exception or combine two passages incorrectly. That’s why the source sits next to the answer. Check details with legal or financial consequences there.
Yes, if the system offers an interface – a ticketing or CRM system, for example. Reading data, preparing entries, triggering workflows: which systems we connect, and with which permissions, is clarified in the project. Only steps defined in advance are carried out, and only after confirmation if you wish.
There are no fixed packages. The effort for development and operation depends mainly on how many sources are connected, whether the chatbot carries out actions and which infrastructure it runs on. We give a reliable estimate once the use case has been clarified: in the requirements description with effort estimate.
Yes, for example questions about plans, products or terms. In that case we recommend stricter rules: only documents approved for customers, clear topic limits and a route to your customer service when the chatbot can’t help.
For IMAP GmbH and the Federal Anti-Discrimination Agency (ADS), we are developing an AI-powered survey chatbot. It captures experiences of discrimination in Germany in a structured way, with a conversation flow that adapts to the answers and a focus on accessibility. It guides people through a survey rather than answering from documents.
Related to knowledge chatbots
Discuss your knowledge chatbot
Tell us which questions and documents keep your team busy.