Ask Prism
Document Analytics Platform with visual citations—click a citation and the PDF scrolls to the exact passage, highlighted with bounding boxes.
Demo: AI-Powered Table Enrichment
Upload PDFs, ask questions, get answers with clickable citations that highlight exact source text
40+
3 Models
Multi-model
~96%
Dual
LangGraph + Supabase + pgvector
Architecture
┌─────────────────────────────────────────────────────────────────┐
│ VERCEL │
│ ┌───────────────┐ ┌───────────────┐ ┌──────────────────┐ │
│ │ Next.js │ │ API Routes │ │ Upstash Redis │ │
│ │ Frontend │ │ (SSE stream) │ │ (Query Cache) │ │
│ └───────────────┘ └───────────────┘ └──────────────────┘ │
└─────────────────────────────────────────────────────────────────┘
│ │
▼ ▼
┌─────────────────┐ ┌─────────────────────────┐
│ LangGraph │ │ Supabase │
│ (Q&A Agent) │ │ PostgreSQL + pgvector │
│ │ │ │
│ routeQuery │ │ - documents │
│ ↓ │◄─────►│ - chunks (embeddings) │
│ retrieve │ │ - conversations │
│ ↓ │ └─────────────────────────┘
│ verify │
│ ↓ │ ┌─────────────────┐
│ generate │──────►│ LangSmith │
└─────────────────┘ │ (Tracing) │
└─────────────────┘Key Features
Visual Citations
Click any citation to scroll the PDF to the exact location with bounding box highlighting.
How: PDF parsing extracts text with position data (bounding boxes), chunks preserve bbox coordinates, and react-pdf-highlighter-extended renders highlights.Multi-Model Verification
Dual LLM verification catches hallucinations before users see them.
Pipeline: GPT-5 Mini (primary) → Claude Haiku 4.5 (verification) → o3 (reconciliation if disagreement)Chain of Thought UI
Transparent AI reasoning with collapsible step-by-step visualization.
Ported to: Storybook Design SystemAI Elements Library
40+ reusable components for building AI interfaces.
Categories: Reasoning, Messages, Verification, Citations, Processing, CodeKey Design Decisions
Why dual parsers?
pdfjs-dist is free but struggles with complex layouts. Reducto handles tables and scans but costs credits. Users choose based on document complexity.
Why multi-model verification?
Single LLMs can hallucinate. Using different models (GPT-5 Mini + Claude Haiku) catches errors. Research shows dual verification achieves ~96% accuracy vs ~85% baseline.
Why bounding boxes?
Page-level citations aren't precise enough. Bounding box extraction enables exact text highlighting, building trust and enabling quick verification.