Problem: An organisation with sensitive documents could not send them to a cloud AI service. They still wanted to ask questions of their own material and get answers drawn from it.

Approach: We deployed a complete local RAG system: ingestion, embedding generation, vector storage, and answer synthesis. Nothing leaves the local hardware, and there is no external API call anywhere in the path.

Technical details:

• Ingestion for PDF, Markdown, and plain text
• Chunking and embedding with local models
• Vector search with configurable top-k retrieval
• Answer synthesis from retrieved context
• Docker Compose deployment for every component
• Configurable chunk sizes, overlap, and retrieval parameters
• API for integration with existing applications

Outcome: Staff query their documents in plain language and the data stays in the building. A document update is reflected in search results within minutes.

Stack: Python, Vector Database, Docker Compose, Local LLM, Embedding Models

Sandboxed plugin ran for projects/rag-pipeline and read “Local RAG Pipeline” through the mediated API.