Short answer
Retrieval-augmented generation answers questions from your own documents, but quality depends on retrieval, permissions and evaluation, not on the model alone. We build RAG and enterprise search that respects access rights, cites sources and is measured against real questions before and after go-live.
On this page
Why most internal search still disappoints
Employees in large organizations spend a lot of time looking for things: the current version of a policy, the last offer to a customer, the technical specification for a product variant, the decision from a meeting two quarters ago. The information exists, spread across SharePoint, file shares, Confluence, ticket systems, ERP attachments and email archives. Keyword search returns long lists of outdated files, and a quick chatbot prototype on top of a document folder answers confidently but cannot say where its answer came from, or answers from a document the user should not see.
RAG, retrieval-augmented generation, fixes that only if the retrieval part is engineered properly. The language model can only be as good as the passages it receives.
What a production RAG system needs
Clean ingestion
Connectors pull content from each source with its metadata: owner, date, version, language, document type and access control lists. Documents are parsed with attention to tables, headings and scanned pages, then split into passages that keep their context. Outdated versions and duplicates are filtered or down-ranked.
Hybrid retrieval
We combine keyword search, which is strong on product codes, names and exact terms, with vector search, which finds passages that mean the same thing in different words. A re-ranking step orders the candidates before the model sees them. Metadata filters narrow the search by department, product, date range or language.
Permissions at query time
Every result is filtered by what the asking user may see in the source system, using the access lists captured at ingestion and refreshed on change. A user should never get an answer built from a document they could not open themselves. We test this with dedicated cases for each permission model.
Answers with sources
Answers cite the passages they are based on with links back to the original document, so staff can verify before acting. When the retrieved material does not support an answer, the system says so instead of guessing.
Measuring quality instead of hoping
We build an evaluation set together with your subject experts: real questions, the documents that should answer them and what a correct answer contains. We then measure retrieval quality (did the right passages come back?) and answer quality (is the answer correct, complete and grounded in the sources?) separately. That split tells us whether to fix chunking, ranking, metadata or prompts. The evaluation runs on every change and on a schedule in production, so a new document structure or model update does not degrade results unnoticed.
Data protection and confidential content
Enterprise content often contains personal data, HR files or customer information. We define with your data protection officer which sources are in scope and which are excluded. Where content must be processed by an external model, Textwash Pro can anonymize text locally before it leaves your environment. Where that is not acceptable, we run retrieval and models on EU-hosted or on-premises infrastructure.
Responsibility split
Our AI Code Engineers own connectors, the ingestion pipeline, retrieval design, permission filtering, the evaluation suite, monitoring and documentation. Your side owns source selection, content ownership, the definition of correct answers for critical topics and the decision to go live for each user group. Content owners remain responsible for keeping their documents current; the system can flag stale or contradictory documents to them.
For organizations that want a long-term partner to run and extend the search under service levels, OnlineAppCreator delivers custom AI development for B2B and public-sector customers with SLAs.
Delivery in practice
- Discovery: sources, user groups, top questions, permission models.
- Pilot on two or three sources with one user group and a first evaluation set.
- Iteration on retrieval quality until the agreed quality bar is reached.
- Rollout to further sources and groups, with SSO and audit logging on the enterprise AI platform.
- Operation with scheduled re-indexing, evaluation runs and usage review.
What to expect
Staff find answers faster and can check them against the source. Questions that used to go to a few experts get answered by the system with a citation, and the experts see what people are asking. Access rights stay intact. And because quality is measured, you know where the system is strong and where content needs work.
More on our AI code engineering service. Related: SharePoint AI knowledge assistant and AI feature development for SaaS.
FAQ
What is RAG in simple terms?
The system first searches your documents for relevant passages and then lets a language model write an answer based only on those passages, with references. The model does not need to be trained on your data.
How do you make sure users only see what they are allowed to see?
Access rights from the source systems are stored with each document and applied at query time for the asking user. We verify this with test cases for each source.
How do you measure whether answers are good?
With an evaluation set of real questions and expected answers created with your experts. Retrieval and answer quality are measured separately and tracked over time.
Can it run without sending data to a US cloud provider?
Yes. We can use EU-hosted models or run models on your own infrastructure, and anonymize text locally where an external model is used.