Sovereign Multimodal RAG

Ask your industrial
documentation.

A self-hosted AI that answers questions over German industrial PDFs (Siemens, Bosch, TRUMPF, KUKA, Festo) in German and English. No OCR. No data sent to OpenAI. EU AI Act compliant by design.

Live Demo View Code

„Geschäftsgeheimnis gehört nicht in OpenAIs Logs."

Trade secrets don't belong in OpenAI's logs.

The Problem

Why German industry needs this

Mittelstand companies have 40+ years of technical knowledge locked inside PDFs. They want AI to unlock it. But the usual tools don't fit.

1 Data must stay on-prem

Engineering drawings and datasheets are IP. They cannot be sent to OpenAI or any US cloud. Lastenheft runs entirely on your own hardware.

2 EU AI Act is active

Industrial AI often counts as high-risk. Every deployment needs risk classification, transparency, and an audit trail. Built in here, not bolted on.

3 Diagrams break OCR

Technical pages are full of tables, schematics, and drawings. Text-only systems lose them. Lastenheft reads the page as an image, directly.

4 Precision matters

When tolerances and certifications are on the line, a hallucinated answer is worse than no answer. Every fact here is cited to a source page.

How it works

Four agents, one answer

You ask a question. A multi-agent pipeline finds the right pages, checks them, and writes a cited answer. You watch every step happen live.

1 · Planner

Understands

Breaks your question down and decides how hard it is.

→

2 · Retriever

Finds pages

ColPali visual search over 900+ pages. No OCR. Works on diagrams.

→

3 · Validator

Checks

Confirms the found pages can actually answer the question.

→

4 · Synthesizer

Answers

Writes the answer with a citation on every fact. Local or API.

The interface

See it in action

Clean, fast, German-industrial. Dark interface built for long reading sessions.

Home. Sidebar history, sovereignty toggle, example queries.

Live trajectory. Watch the four agents run, with timing.

Cited answer. Provider, cost, latency. Click [1] to see the source.

History. Past queries grouped by day. Delete any of them (GDPR).

Compliance, Article 6. Risk classification with documented reasoning.

Compliance, Article 13. A real audit log of every AI call.

Mobile. Fully responsive.

Drawer. History on small screens.

Measured, not claimed

The numbers

Evaluated on 108 held-out questions over 909 pages. The fine-tuned reranker beats the off-the-shelf one by a clear margin.

70.4%

Hit@1

correct page ranked first

0.758

MRR

+8 pts from fine-tuning

0.752

Faithfulness

answers grounded in sources

Cost / query

in local sovereign mode

Retrieval strategy	MRR	Hit@1	Hit@10	nDCG@10
ColPali visual search (baseline)	0.341	20.4%	61.1%	0.395
+ BGE reranker (off-the-shelf)	0.707	62.0%	85.2%	0.743
+ BGE reranker (my LoRA fine-tune)	0.758	70.4%	85.2%	0.781

Compliance

EU AI Act, designed in

Not a marketing slide. Every requirement maps to working code and a real database table.

Article 6: self-classified risk category with documented rationale.
Article 13: every answer shows which AI made it, with citations.
Article 14: human oversight, with a full audit log you can review.
Article 10: data lineage from answer back to source page.
GDPR Art. 17: right to erasure. Delete any query or session.
GDPR Art. 25: privacy by design, data residency configurable.

Under the hood

Built end to end

Real ML, real engineering, real deployment. Every layer chosen for sovereignty and precision.

ColPali v1.3 · visual retrieval, no OCR BGE reranker + LoRA · fine-tuned LangGraph · multi-agent Qwen3 4B · local LLM via Ollama Claude Sonnet · optional API escalation FastAPI · Python backend Next.js 16 · React frontend Postgres + pgvector · vector DB Langfuse · observability Docker Compose · one-command deploy

Open source · MIT

Explore the project

Full source code, reproducible eval, training scripts, and a one-command Docker setup. Clone it, run it on your own hardware, point it at your own PDFs.