AI knowledge systems
AI knowledge search for grown SharePoint documentation
A document base that had grown over years was made semantically searchable. Employees could ask questions in natural language and received fitting answers with traceable sources.
- Area
- AI knowledge systems
- Technologies
- SharePoint, Python, Gemini API, semantic search
- Project type
- Prototype and technical implementation
Starting point
-
01
Over several years, different teams had stored operations documentation, process descriptions, technical documents and further internal knowledge in SharePoint.
-
02
The information was there in principle. The filing structures had grown historically, though, and were therefore not always easy to follow.
-
03
Employees often already had to know in which area or folder they needed to search. Similar content sometimes sat in several places. New team members depended more on experienced colleagues just to find relevant documentation at all.
-
04
It was also hard to see which content existed more than once or was spread across different filing areas.
Goal
The existing documents were not supposed to be moved into a new system. SharePoint was to remain the central store.
What was needed was a way to make the existing knowledge base accessible through natural language.
Typical questions were meant to include:
- Where do I find the operations documentation for system X?
- Which documents exist for process Y?
- Which documents cover a similar subject?
- Where might duplicate or redundant content sit?
- Which documents are especially relevant for this topic?
It mattered that answers were not produced in isolation by a language model. The underlying sources were to stay traceable at all times.
Architecture
- 01 SharePoint
- 02 Document import with Python
- 03 Extract text and metadata
- 04 Prepare and index content
- 05 Semantic search and retrieval
- 06 Gemini API
- 07 Answer with sources
SharePoint remained the leading document system.
A Python service took over reading and preparing the relevant content. Matching document passages were found for a user question and then passed to the language model together with the question.
For the AI processing, the Gemini API that was already in place and approved was used.
The architecture was deliberately built so the language model stays swappable. Depending on infrastructure and data protection requirements, a locally run model via Ollama could be connected, for example.
Implementation
Among other things, the following parts were implemented in the project:
SharePoint connection
Relevant documents and metadata were taken from existing filing structures.
Document preparation
Larger content was processed and split into suitable sections for later search.
Semantic search
Matching content was not found only via file names or tags, but by its meaning.
Gemini integration
The information found was used as context for answering the user questions.
Source referencing
Answers were output together with the underlying documents.
Analysis of the document base
In addition, first approaches were used to make similar or possibly redundantly stored content visible.
Result
Searching for internal knowledge could be simplified clearly.
Faster research
Employees no longer had to search several SharePoint areas and folder structures by hand.
Traceable answers
The relevant sources were shown with the generated answers. Information could therefore be checked immediately.
Less dependence on filing knowledge
Users no longer had to know exactly where a piece of information was originally stored.
More transparency
Thematically similar or scattered content became easier to see and could be used as a basis for later clean-up.
Before and after
Search via folders and filing structures
Search via natural language
The filing location has to be roughly known
Content decides relevance
Check several documents by hand
Answer with matching sources
Distributed content hard to see
Thematically similar content visible
Knowledge partly sits with experienced staff
Documented knowledge becomes easier to access
Why this solution
- No extra AI platform was introduced for the project.
- Because the Gemini API was already in place and organisationally approved, the existing infrastructure could be used.
- At the same time the solution was not locked to a single vendor. The language model remains a swappable part of the architecture.
- The same basic solution can therefore be implemented with a cloud API or a locally run model, depending on the constraints.
Room to grow
Further functions can be built on the existing architecture.
Permission-aware search
Existing SharePoint permissions can also be taken into account in the knowledge search.
Detection of redundant content
Similar or multiply maintained documentation can be identified systematically.
Detection of outdated information
Metadata, versions and overlapping content can provide hints of outdated documents.
Spot knowledge gaps
Recurring questions without a sufficient documentation basis can be made visible.
Connect further data sources
Besides SharePoint, wikis, ticket systems or other internal knowledge sources can be added, for example.
A similar use case?
If existing knowledge is documented but staff find it only with difficulty, or grown filing structures become increasingly hard to overview, an AI-supported knowledge search can be a sound approach.