Demo 2
RAG Demo (Technical)
This demo answers questions over a fixed corpus of real NIST and NASA documents. It translates when needed, retrieves the matching passages and answers with page citations – ask in German, source in English, answer in your language.
The document base
We don't host the PDFs ourselves but link the original sources (public domain, NIST/NASA).
Topics & sensible questions
Clicking copies the question. Paste it into the demo below and send.
Expected: the core focus areas of the AI profile, citing NIST IR 8596.
Expected: a short definition plus attack categories, citing NIST AI 100-2e2025.
Expected: the foundational activities from NIST IR 8259r1.
Expected: key subsystems, citing the NASA SoA report.
Live demo
The demo runs on a separate subdomain and loads only after you click.
Security & privacy
- - The corpus consists solely of public-domain documents (NIST/NASA).
- - Every answer cites its source and location – traceable, not guessed.
- - Abuse protection via request and input-length limits.
- - Inputs are not stored as training data.
- - Please do not enter confidential or personal data.
- - Answers may contain errors – when in doubt, check the linked original source.
Technical details
- - Adaptive/corrective RAG pipeline orchestrated with LangGraph (retrieve → grade → optional self-correction → generate).
- - Cross-lingual embeddings (qwen3-embedding-8b, 4096 dims): German questions retrieve English passages without translation.
- - Vector search with Qdrant over ~4,000 chunks; exact and semantic caching for fast repeats.
- - Two modes: “fast” (1 LLM call) and “showcase” (full graph with grading & self-correction).
- - Answer model openai/gpt-oss-120b via OpenRouter; hosted on dedicated infrastructure (Hetzner, EU).