The Challenge
An established consulting firm with over 15 years of history faced a classic knowledge problem: decades of project experience, methodological expertise, and client insights were trapped in email silos, file servers, and the heads of individual consultants. New hires needed months to get up to speed. Senior consultants spent hours searching for old project documents.
The existing file server search only matched filenames — not content. If you didn't know where a document was stored, you couldn't find it. At the same time, the company had strict data protection requirements: client data was not allowed to leave their own infrastructure.
Our Approach
Blueprint Phase: Knowledge Mapping
We mapped existing knowledge sources: email archives (Exchange), file servers (SMB shares), a wiki (Confluence), and personal notes in OneNote. For each source, we defined connectors and indexing strategies.
Brain Phase: RAG Architecture
The central design decision: fully on-premise. No cloud LLM, no external APIs. We selected an open-source language model running on the company's own hardware. The vector database (pgvector) was integrated into the existing PostgreSQL infrastructure.
Hands Phase: Incremental Indexing
Indexing was done source by source. Each new data source was connected, tested, and validated before adding the next. This allowed us to optimize retrieval quality per source.
Architecture
Document Connectors
Specialized connectors for each data source: Exchange connector for emails and attachments, SMB connector for file server documents, Confluence API connector for wiki pages. Each connector extracts text, metadata, and relationships.
Vector Database
PostgreSQL with the pgvector extension stores embedding vectors directly alongside source documents. Hybrid search combines semantic similarity with keyword matching for optimal results.
RAG Pipeline
User queries are processed in real time: embedding generation, vector search, result re-ranking, and answer generation with source citations. The language model only sees relevant document excerpts — never the entire knowledge base.
Access Control
The system respects existing permissions: users only see documents they have access to in the source system. Permission checks happen in real time with every query.
Results
- Access in under 1 second — 15 years of company knowledge instantly searchable
- Fully on-premise — no data leaves the company infrastructure
- GDPR-compliant — open-source model without external API calls
- Source citations — every answer references the original documents
- Existing permissions — access rights are automatically respected
- Faster onboarding — new consultants find relevant knowledge from day one
Facing a Similar Challenge?
Your company knowledge is scattered across silos or you need a GDPR-compliant knowledge solution? We build RAG systems on your own infrastructure. Talk to us or learn more about our data infrastructure services.
