RAGnar Knowledge Processor

Organisations hold plenty of useful documentation, but it is hard to use day to day. Instructions sit in PDFs, policies in DOCX files, procedures in shared folders, and people still ask where things are.

A generic chatbot does not solve this. Without access to sources it makes things up. If a large PDF is dropped in without preparation, retrieval returns oversized chunks and the model tries to answer from incomplete context.

Two connected layers. RAGnar is a self-hosted local RAG platform for building assistants over private documentation. RAGnar Knowledge Processor is a Python preparation pipeline that cleans PDF and DOCX files, splits them by topic and saves them as small Markdown knowledge files.

RAGnar lets you choose the model, configure retrieval, work with local embeddings and a Qdrant vector database, and display sources. The processor prepares the input so the assistant never has to work on large, noisy documents.

1. A PDF or DOCX enters the Knowledge Processor.
2. The pipeline extracts and normalises the text and detects headings and topic sections.
3. Content is saved as small Markdown files with titles, categories, tags and source references.
4. RAGnar ingests those files, generates embeddings and stores them in Qdrant.
5. A user creates an assistant and tunes chunk size, top-k, similarity threshold and temperature.
6. Each answer shows the documents, chunks and source snippets, so the context can be checked.

A Docker Compose environment. Open WebUI is the chat interface, Ollama runs local models, Qdrant is the vector store, FastAPI manages the RAGnar endpoints and configuration, and NGINX handles routing between the interface and the API.

The emphasis is transparency: the assistant should not only return an answer, it should show which document and which part of the text was used.

• Runs locally, so documents do not have to leave the organisation's infrastructure.
• Each assistant has its own knowledge and retrieval settings, which limits what one assistant can reach.
• Source display makes unsupported answers easier to spot.
• Next step for this system: adversarial testing of the retrieval layer for prompt injection through documents and for leakage between assistants.

RAGnar is a base for several internal assistants: HR procedures, IT documentation, Linux guides, networking scripts, security policies or onboarding material. Instead of one general assistant, each can have its own documentation and retrieval settings.

The Knowledge Processor solves a step many AI projects skip: preparing knowledge before the chat. Smaller, cleaner Markdown sources give steadier retrieval, fewer hallucinations and clearer citations.

The biggest lesson was not that a bigger model is needed. It is that a better input is needed. Local models can be useful when documentation is well prepared, chunks are meaningful and the prompt is restrictive enough.

No single part is enough on its own. Infrastructure, the document pipeline, vector search, prompt rules, source citations and constant testing of answers have to work together.