VaultIQ: Governed Document Intelligence

Enterprises sit on vast archives of image-based documents — scanned reports, government publications, multilingual PDFs — that are impossible to query at scale. Existing tools either require clean text or work in a single language. Neither assumption holds in the real world. VaultIQ solves this with a Visual RAG pipeline: instead of extracting and indexing text, it uses PageIndex to build a visual reasoning graph directly from the PDF pages. When a user asks a question, PageIndex identifies which pages are relevant — then Llama 4 Scout (via Groq) reads those raw page images and generates a grounded answer with explicit page citations. No OCR errors. No layout destruction. Tables, charts, and Arabic typography are all handled natively. On top of the RAG pipeline, Lobster Trap acts as a governance proxy — sitting between the application and the LLM backend to enforce PII detection, firewall rules, rate limiting, and full JSON audit logging. Prompt injection attempts are blocked outright. High-risk queries are flagged for human review. Every interaction is logged. Enterprise-grade auditability, out of the box. POC: Arabic and English querying over the Madinah Tranquil Livable City 2024 report, achieving 93% answer accuracy vs. 81% for a text-based baseline — a +12 percentage point improvement driven entirely by reading the pages visually. Stack: PageIndex · Groq (Llama 4 Scout) · Lobster Trap · Streamlit · Python