AI PoC Speedboats
With AI PoC Speedboats, T-NEX evaluates a limited AI use case: criteria agreed before starting, checks against examples and a decision on the next step.
Explore serviceKnowledge assistants for regulated markets
Find expert knowledge without searching through documents.
A knowledge assistant makes approved policies and procedural knowledge accessible to your team. Answers stay linked to their sources. We align search, permissions and specialist follow-up questions with your content and intended use.
Cobalt · Knowledge assistant
Find approved knowledge with its source.
Find a relevant answer and open its source. Maintained questions, answers and synonyms connect mobile search with knowledge administration.
Maintain knowledge entries
Choose a question
The mobile start page offers suggested questions about leave and onboarding alongside recent conversations. Selecting a question opens the knowledge assistant.
A specific question leads into the chat.
Enlarge viewRead the answer and source
A maintained answer explains leave requests and carrying leave forward. The visible source reference names Leave Policy v3.2, section 4.
The source reference stays visible with the answer.
Enlarge viewMaintain knowledge entries
The administration view places answers, categories, source documents and search terms side by side. Entries can be added and edited directly.
Answer text and its source are maintained together.
Enlarge view In daily work
When is a maintained answer library useful?
For business teams handling recurring internal questions about policies and procedures.
An institutional knowledge assistant can start with an editorially maintained library. Each entry contains a question or title, an approved answer, a category and its source document. Section references and synonyms help users find the answer despite differences in wording.
This approach suits recurring work questions with an authoritative answer. The assistant finds the matching entry and displays its maintained text. Where no sufficiently relevant content exists, the question is returned for clarification.
Functions
How the application supports your team.
01
Retrieve approved answers
For clearly defined questions, the assistant can return maintained answers with a recorded source.
02
Develop a generative RAG layer
Where a question requires several passages, a dedicated answer workflow is built and checked against business test cases.
03
Assign knowledge ownership
Content has a responsible editorial owner. Source references and updates are part of operating the system.
04
Handle unanswered questions
Unanswered questions are handled explicitly and directed to business review.
Overview
When is generative RAG useful?
Retrieval-Augmented Generation, or RAG, combines finding relevant passages with composing a new answer. This can help when a question requires information from several documents or sections. The retrieved source and generated statement are treated as separate review objects.
For this extension, we define permitted documents, meaningful passage boundaries and the conditions for answering. Business test cases check whether the selected passages actually support the statement. Quality also depends on how document editions, exceptions and conflicting sources are maintained.
Overview
How is knowledge kept current?
Knowledge maintenance needs a dedicated working environment. Responsible departments edit entries, categories, source references and search terms. Superseded content is corrected or removed. Policy changes require an assigned owner to review affected answers.
Unanswered questions identify concrete editorial work. Recurring questions may reveal an unclear passage or an undocumented process. Usage analysis supports improving the knowledge base; the scope of retained conversation data is defined before operation.
Overview
Which boundaries and handovers does the assistant need?
An internal assistant needs clear content-access rules. Access to the application must not automatically provide access to every document. Intended roles, datasets and user groups are therefore checked against the actual use case.
For missing, conflicting or unapproved information, we define a clear refusal and a reachable clarification path. Handover can include the question and permitted context. Authority to make a decision or change a policy remains separate.
Overview
How is quality checked before rollout?
We begin with real recurring questions from a limited subject area. These form a test set containing expected answers, supporting sources and cases that should receive no answer. Abbreviations, ambiguous wording and missing context are included.
Acceptance considers answer correctness, source support, access boundaries and handling of knowledge gaps. It also defines how new document versions and changes to the answering approach are checked again. A successful single example is insufficient.
The workflow
How we introduce the solution.
- 01
Select recurring questions and permitted sources.
- 02
Choose retrieval or a generative RAG layer.
- 03
Set up knowledge maintenance and access permissions.
- 04
Review answers with business owners and test unanswered cases.
What needs to be agreed before use
RAG does not eliminate incorrect answers. If sources are outdated or do not cover a question, the system needs to withhold an answer and provide access to a business contact.
FAQ
Questions about the solution
Is every chatbot a generative RAG system?
No. A retrieval assistant can display stored answers unchanged. Generative RAG is a separate extension with additional business validation.
Which content makes a useful first use case?
Approved policies, operating procedures and frequent internal questions with clear ownership are good starting points. A limited, maintained collection is easier to validate than an unstructured archive.
Can the assistant handle abbreviations and different terminology?
Synonyms and search terms can be maintained in an answer library. For a generative solution, realistic wording variants are included in acceptance testing.
What if the documents do not contain an answer?
A clear refusal and business clarification route are designed for that case. A plausible-looking citation must not conceal an unsupported statement.
Who owns the answers?
The responsible business departments own the content. T-NEX implements the search, editing and review processes and helps assess the answering approach.
Related options
You may also be interested in these.
Evaluation and optimisation
T-NEX reviews applications and infrastructure for regulated markets, with an assessment, technical evaluation and a plan for optimisation or migration.
Explore serviceRegulated software development
Develop applications on the T-NEX platform with defined data access, testable acceptance criteria and agreed rights, handover and operational responsibilities.
Explore serviceWhich task would you like to solve next?
Bring a concrete task. Together, we will define what the application needs to do.
Discuss your knowledge sources and use case