Dome Document Intelligence
Extracts structured data from any business document — with per-field confidence scores, governance validation, and a full audit trail — without storing the document.
Covers the Orchestrate & Model phases of the DOME method. Part of the Dome portfolio. For cross-cutting Dome context see dome-docs.
What it does
Upload a PDF, image, or spreadsheet and the tool:
- Auto-detects the document type (invoice, purchase order, trade confirmation, contract, etc.) — no templates or configuration needed
- Extracts all significant fields with per-field confidence scores
- Validates the extraction against a configurable set of governance rules
- Logs an audit event (metadata and hash only — document content is never stored)
- Flags whether human review is required
The key differentiator within the Dome portfolio: this tool is designed to run with a locally deployed AI (Ollama), meaning no document content ever leaves the client's infrastructure. Cloud (Claude API) and Azure OpenAI modes are also supported via a provider switch.
Status
| Layer | Status |
|---|
| Backend (FastAPI) | Complete |
| Frontend (Next.js) | Complete |
| Claude provider | Working |
| Ollama provider | Stubbed — in progress |
| Azure OpenAI provider | Stubbed |
Evaluation
Extraction quality and the governance rules engine are measured by an auditable, four-layer
evaluation harness in backend/eval/ — see EVAL.md for the full
methodology and a sample report.
- Rules engine — all 16 governance rules are unit-tested (CI-gated, deterministic).
- Extraction — scored against a committed golden set with type-aware normalization: field
precision/recall/F1, value-match, doc-type accuracy, end-to-end human-in-loop agreement.
- Confidence calibration — a reliability table + ECE that checks the model's stated
confidence against its actual accuracy.
- Validated LLM judge — checked against ground truth before it is trusted for fuzzy fields;
each live run logs one governance
eval_judgment event.
cd backend
make eval # live: re-run the model + validated judge → eval_report.{json,md}
make eval-score # offline: score the committed recordings (no network)
CI runs the deterministic layers against recorded outputs and never calls a live model.
Quick start
Requirements
- Python 3.12
- Node.js 20+
- Ollama (optional — for local AI mode)
Setup
# Clone
git clone https://github.com/Dome-Layer/dome-document-intelligence.git
cd dome-document-intelligence
# Backend
cd backend
python -m venv venv
source venv/bin/activate # Windows: venv\Scripts\activate
pip install -r requirements.txt
# Environment
cp .env.example .env
# Edit .env with your values — see Environment variables below
Run locally
# Backend (from backend/)
uvicorn app.main:app --reload --port 8000
# Frontend (from frontend/)
npm install
npm run dev
Backend API: http://localhost:8000. Health check: GET /api/health.
Frontend: http://localhost:3000.
Environment variables
Create backend/.env from .env.example. Required variables:
# LLM — choose one mode
LLM_PROVIDER=claude # claude | azure_openai | ollama
# Claude (cloud demo)
ANTHROPIC_API_KEY=sk-ant-...
CLAUDE_MODEL=claude-sonnet-4-6 # optional — override the Claude model
# Azure OpenAI (client tenant — not yet implemented)
AZURE_OPENAI_ENDPOINT=
AZURE_OPENAI_KEY=
AZURE_OPENAI_DEPLOYMENT=
# Ollama (local / air-gapped — not yet implemented)
OLLAMA_URL=http://localhost:11434
# Supabase
SUPABASE_URL=
SUPABASE_ANON_KEY=
SUPABASE_SERVICE_ROLE_KEY=
# App
ENVIRONMENT=development # development | production
CORS_ORIGINS=http://localhost:3000
MAX_FILE_SIZE_MB=20
Never commit real secrets. backend/.env is gitignored; .env.example is not.
API endpoints
| Method | Path | Description |
|---|
POST | /api/extract | Upload a document and run extraction |
GET | /api/rules | List governance rules for the authenticated user |
PATCH | /api/rules/{rule_id} | Toggle a rule enabled/disabled |
GET | /api/audit | Paginated governance event log |
GET | /api/audit/{event_id} | Single audit event detail |
GET | /api/health | Health check |
All endpoints except /api/health require a Supabase bearer token. Authentication uses Supabase magic link, triggered from the DOME website login page.
Example extraction call
curl -X POST http://localhost:8000/api/extract \
-H "Authorization: Bearer <supabase-token>" \
-F "[email protected]"
Database setup
Run the canonical DOME governance schema — dome-docs/templates/governance_schema.sql — in your Supabase SQL editor to create the required tables (governance_rules, governance_events) and indexes. For in-place upgrades to an existing deployment, apply the numbered files in dome-docs/migrations/. Governance rules for each user are seeded automatically on first GET /api/rules.
Deployment
Cloud demo
- Backend: Railway (Dockerfile deploy, EU region)
- Frontend: Vercel (EU region) — not yet deployed
Client Azure tenant
Azure OpenAI provider not yet implemented. See TOOL_CONTEXT.md for planned architecture.
Air-gapped / local AI
Ollama provider not yet implemented. Target models: Llama 3.2 Vision, Qwen2-VL. See TOOL_CONTEXT.md for planned architecture.
Repository structure
dome-document-intelligence/
├── README.md This file
├── EVAL.md Evaluation methodology + sample report
├── LICENSE
├── SECURITY.md
├── backend/ FastAPI service
│ ├── Makefile eval / eval-score / test / lint targets
│ ├── eval/ Evaluation harness (golden set, scorer, calibration, judge)
│ ├── app/
│ │ ├── main.py App entry point, lifespan, CORS
│ │ ├── api/ Route handlers (extract, rules, audit)
│ │ ├── core/ Config, DB client, logging
│ │ ├── models/ Pydantic schemas
│ │ ├── providers/ LLM provider implementations
│ │ └── services/ Ingest, extraction, validation, governance
│ ├── supabase_schema.sql Pointer stub — canonical schema lives in dome-docs/templates/
│ ├── Dockerfile
│ ├── railway.toml
│ └── requirements.txt
└── frontend/ Next.js application
├── app/ Pages (upload, result, rules, audit)
├── components/ UI components (upload, result, layout, auth)
├── context/ Auth context
├── lib/ API client, auth helpers, types
└── styles/
Related
