SecureAgentRAGBYOK demo

9-node LangGraph · RBAC at vector DB · NLI faithfulness gate · SHA-256 audit chain · session

github
Using the owner's throttled key (10 requests / hour / IP). Paste your own Groq / OpenAI / Anthropic key for unlimited use — it lives only in your browser's localStorage and never reaches our database.
Persona:

Privacy-first multi-agent RAG demo.

A curated set of documents is indexed in a Qdrant vector store with an RBAC payload filter. Pick a persona above and ask a question — chunks you are not authorized to see are physically not returned, regardless of similarity score. The 9-node LangGraph below handles routing, prompt-injection guardrails, retrieval, grading, synthesis, and an NLI faithfulness gate that flags any sentence the cited chunk does not entail.

👉 Tap 📚 Docs in the header to see every document in the knowledge base and exactly which ones this persona can read — or 📎 Upload your own.

🔓 Privacy note: this public demo has no local model, so all inference — including content classified high sensitivity — runs on Groq cloud. The sensitivity: badge on each answer tells you when that happened. Self-host with a local Ollama for the “HIGH never leaves local” guarantee — how & why.

Try one of these — tuned for engineer:

افهم عقدك — اسأل بالعربية

نفس النظام يجيب بالعربية ويستشهد بالمصدر من مستندات توضيحية مصرية (عقد إيجار، قانون العمل، التسجيل الضريبي). جرّب سؤالًا: