Model implementation of the internal knowledge base AI based on MAG, which indexes documents, procedures and instructions, and then responds to employees on the basis of company sources.
A model customer is a company with over 20 employees, several hundred documents, dozens of procedures, multiple locations and distributed materials between SharePoint, OneDrive, Google Drive, Teams, PDF, Word, Confluence, Wiki, email and network folders. Each department has its own instructions and new employees often ask for the same rules, documents and processes.
The employees did not know which version of the document is current, where the procedure is, what the process of complaint, shopping, onboarding or system handling looks like. Experts were constantly distracted from work, and the classic search engine did not understand questions in the natural language and did not show an answer with sources.
Every day, employees look for procedures, instructions, templates, complaint policies, onboarding information, security policies and answers about systems. RAG does not create new knowledge. It helps find existing knowledge, indicate sources and limit repetitive questions to experts.
Indicative values. Each indicator is an example value that requires confirmation during pre-implementation analysis, based on real documents, user questions and access rules.
Each department has its own materials, folders and shortcuts. New employees ask about the basics, and experts are constantly distracted from their work. The knowledge exists, but it is scattered across SharePoint, OneDrive, Google Drive, Teams, PDF, Word, Confluence, Wiki, email, and network folders.
An internal AI knowledge base makes the most sense when a company has a lot of material, but employees still ask people because they don't trust the folder structure or don't know what to name the document they're looking for.
Documents exist but are scattered across folders, emails, Teams, SharePoint, OneDrive and local copies.
Employees waste time searching instead of completing the task.The same procedure appears in several files and the user does not know which version is current.
The company risks working on outdated principles and making incorrect operational decisions.Instructions following process changes are not always published in one place and quickly lose credibility.
The team asks experts more often because they don't trust the documentation.The classic search engine does not understand questions in natural language and requires knowledge of file names.
New and less technical users are quicker to give up searching on their own.Onboarding raises the same questions about procedures, tools, leaves, systems and responsibilities.
Experts are distracted from their work and repeat the same answers.Knowledge is in the heads of several people, so Teams and the phone become an informal search engine.
The hidden cost increases with the number of employees and locations.The system does not find an answer when the user asks a question with words other than the document name.
Data is available but virtually invisible to employees.The new employee must learn the folder structure, departments, document names and informal abbreviations.
Onboarding becomes dependent on caregiver availability.The folders have been created over the years by various departments and no longer reflect the current process.
Company knowledge is difficult to maintain and organize.It is unknown who changed the document, when it was indexed, and which replies used which version.
It's harder to audit AI responses and maintain trust in the system.The employee first asked in Teams, then sent an e-mail, called an expert or looked for a PDF in folders. The answer often went to only one person, so the same problem returned after a few days.
We do not train the model from scratch and do not ask the team to copy documents to the new system. We create a document index, embeddings, a semantic search engine, source control, an administrator panel, response history and permissions consistent with the organization.
RAG requires control of the entire process: sources, OCR, chunking, embeddings, index, responses, feedback and maintenance by the administrator.
A file from SharePoint, OneDrive, Google Drive, PDF, Word or a network folder enters the indexing process.
Role: visit • status: syncA file from SharePoint, OneDrive, Google Drive, PDF, Word or a network folder enters the indexing process.
AI finds the current purchasing instruction, describes the acceptance stages and shows the source document.
The employee asks about security rules, and the system indicates the current document and the appropriate fragment.
AI finds instructions for a specific system function ERP and provides steps without searching through folders.
The system finds the correct contract template, version and usage rules for a given customer type.
AI indicates the current procedure, document owner and audit context.
The employee receives the complaint handling process, deadlines, roles and links to forms.
The knowledge base answers questions about leaves, benefits, regulations and onboarding.
AI shows instructions for VPN, MFA, password reset or hardware report.
The operator finds the job manual, quality rules and required checklists.
The system describes the procedure for accepting, complaining about a delivery or issuing a material.
AI indicates invoice circulation instructions, field descriptions and responsible persons.
User finds current policies, reporting procedures and KPI definitions.
The calculator helps estimate the scale of the information problem. It's a starting point for a conversation, not a guarantee of results.
Example value requiring confirmation during pre-implementation analysis. The result depends on the number of questions, quality of documents, authorizations, freshness of sources and the way the team works.
SharePoint • OneDrive • Google Drive • Confluence • PDF • Word • SQL
OCR • chunking • metadane • wersje • walidacja
embedding • pgvector • Qdrant • Pinecone • Weaviate
LLM • RAG • cytowania • guardrails • feedback
Next.js • React • Teams • API • admin panel
Redis • PostgreSQL • monitoring • backup • audit logs
Folder integration alone is not enough. Each source should have an owner, access rules, news status, and a way to update the index.
RAG for company requires permission control at the user, department, document, and source levels. Responses must be consistent with access, and logs should show what fragments the system used.
We inventory sources of knowledge, document types, owners, versions, access levels and frequently asked questions.
At this stage, we also determine which documents can be indexed, which require exclusion and who is responsible for keeping the knowledge up to date.
We design the RAG architecture, roles, permissions, metadata schema, response format, citation rules and scope of MVP.
It is important to determine whether the system is to function only as a search engine or also as an assistant answering employees' questions.
We prepare the processing of scanned documents, PDFs and files without a text layer.
OCR should be tested on real documents, because the quality of scans, tables and stamps may affect the accuracy of the answer.
We divide documents into fragments and create embeddings that enable semantic search.
We select the chunking strategy, embedding model, metadata and the method of updating the index after changing documents.
We connect SharePoint, OneDrive, Google Drive, Confluence, folders, databases or API to the indexing layer.
We check API limits, permissions, synchronization, errors and how to respond to deleted or changed files.
We launch the first version for a selected department, document type or process, for example HR, IT, ISO or sales.
MVP should contain test questions, source citations, a feedback panel and an administrative view.
We test the accuracy of answers, citations, permissions, prompt injection, missing documents and out-of-scope questions.
Only after testing is it worth expanding the scope of documents and making the knowledge base available to a larger group of employees.
After launch, we monitor questions, feedback, knowledge gaps, currency of documents and quality of sources.
The development includes subsequent departments, Teams, Slack, voice search, translations, summaries and document update workflow.
Each indicator is an example value requiring confirmation during pre-implementation analysis.
| Area | Before | After MVP | Measurement method |
|---|---|---|---|
| Information found time | Searching folders and asking experts | Natural language question and answer with source | median time to response |
| Questions for experts | Experts answer repetitive questions | AI supports low-risk questions | number of questions in Teams / email |
| Onboarding | New people ask their supervisor about the basics | The knowledge base leads to procedures and instructions | implementation time for independence |
| Number of documents | Indexing control missing | Source, version and owner status | number of active sources |
| Searches | No data on user needs | Frequently asked questions dashboard | number of searches and topics |
| Accuracy of the answer | No measurement | Feedback and response tests | user rating / review rate |
| Up-to-date knowledge | No one knows what's old | Alerts and document status | number of documents to review |
| Sources | Answer without a trace | Cited excerpts and links | reply share with sources |
Data, KPIs and scenarios are examples. They require analysis of documents, authorizations, sources, quality OCR, current processes and security requirements of a specific organization.
In the first conversation, we will check the sources of knowledge, document types, permissions, frequently asked questions and the safe scope MVP for the first department or process.
In professional implementation, it is necessary to clearly distinguish the use of documents as a context from model training. In RAG architecture documents are indexed, divided into fragments and used to search for answers, but this does not necessarily mean training a public model on company data. The model receives selected fragments as a context of inquiry and prepares an answer on this basis. The decision on where the data are processed, how they are stored and whether they can be used for training depends on the chosen supplier, configuration and safety requirements of the company.
It depends on the architecture. You can design a cloud, hybrid or private solution. In the cloud variant some of the queries can be processed by external APIs, but you still need to limit the data range, use sensitive information masking and control retention. In the private option or on-premise most of the processing can work in the client environment. Before implementation, you need to set legal, technical and organizational requirements and check which documents can be indexed.
Yes, but such a variant requires more infrastructure work. On-premise can include local database, local vector index, private API, own application server, monitoring, backup and AI model run locally or in private cloud. This model makes sense with high security requirements, sensitive data or industry restrictions. However, you need to take into account the cost of maintaining, updating, scaling, testing, GPU resources or integration with a private AI provider.
Indexing starts with downloading the document from a source, for example SharePoint, OneDrive, Google Drive, folder or API. Then the system pulls out text, executes OCR, if the document is a scan, divides the content into fragments, adds metadata and creates embeddingi. Fragments go to the vector database together with information about source, version, owner and permissions. After changing the document, the index should be updated to build on the latest content.
In a well-designed response system should be limited to approved corporate sources or clearly means that there is no document confirming the answer. The language model may have general knowledge, but the system rules should force answers based on documents, procedures and company instructions. If there is no source, the AI should say that no information has been found, and suggest contact with the process owner or report a gap in the knowledge base.
The system should detect changes in documents, delete old versions from the index or mark them as outdated, create new embeddings and update metadata. It is important to indicate the owner of the document and the review date. Thus, the administrator sees which sources are active, which require updates and which responses could be used by the previous version of the document.
Yes. PDFs are one of the typical sources of knowledge in companies, but you have to distinguish PDF with a text layer from a scan. If the file has text, the system can download and divide it into fragments. If it is a scan, you need OCR. OCR quality depends on the quality of the scan, layout of the document, tables, stamps, signatures and language. Therefore, it is worth testing a representative sample of documents before implementation to check which PDFs are suitable for automatic indexing.
Yes. SharePoint is one of the most common sources of documents for companies using Microsoft 365. Integration usually requires Microsoft Graph API, Entra ID applications, appropriate permissions and arrangements which document libraries are to be indexed. Key is to retain permissions: the user should only receive a response from documents to which he has access. Therefore, implementation should include synchronization of sources, metadata, versions and access control.
Yes. Teams can be an interface for asking questions, but it is often worth connecting it to the web application and administrator panel. The employee can ask bot in Teams, and the system responds on the basis of the document index and shows the sources. However, the administrator needs a place to manage sources, feedback, documents to review and logs. Teams is therefore a great channel of use, but does not replace the entire operating layer of the RAG.
Yes, and this is one of the most important elements of implementation. The system should take into account roles, groups, departments, document permissions and levels of confidentiality. The AI reply cannot reveal fragments of a document that you should not see. In practice, this means filtering the search results before passing them to the model and logging which sources have been used. Document-level security should be designed from the beginning and not added at the end.
Yes, this is one of the greatest advantages of the MAG. The answer should include a link to the document, the name of the source, the version or a fragment from which the system was used. This allows the user to check where the information comes from and the administrator can audit the answers. No sources should be treated as a warning signal. Then the system should admit that he did not find a confirmation, instead of generating an answer without the basis in the company documents.
The versioning should keep information about the current version of the document, the index date, the owner, the status of the review and the changes in the source. If the document is replaced by a new version, the index should be updated and the older parts should be deleted or marked as archival. This is important especially for ISO procedures, security policies, technical instructions and HR documents where an outdated response may lead to a malfunction.
It is best to measure the time of finding information, number of repetitive questions to experts, onboarding time, number of searches, relevance of answers, number of documents to update and share of answers with sources. It is worth to start by measuring the current process: how much time employees are looking for information, how often they ask others and which departments are most loaded. Only then can you compare the result after MVP. ROI should not be assumed in advance, only measured on data.
Time depends on the number of sources, quality of documents, permission requirements, OCR, number of departments and MVP range. Simple piloting for one department and selected documents is much shorter than the implementation covering many sources, Teams, SSO, document-level security, admin panel and dashboard. The safest way to start is by auditing documents, lists of test questions and MVPs for one process, and only after the tests develop further sources and departments.
Yes, but each format requires a different way of extracting content. Word is usually easy to process, Excel requires the identification of sheets, tables and context data, and PowerPoint often contains shortcut information, graphics and slides. The system should store file metadata and source information so that the answer can be checked. In case of complex files, it is worth to prepare rules which fragments are to be indexed.
SmartCodeIT can help you design your process, prepare MVP and implement an application tailored to your organization. We start with a short conversation about data, tools, exceptions and business purpose.