Generic Extractor

Config-driven hierarchical document extraction server. Uploads a PDF, runs OCR via Docling, then uses an LLM (Gemini 3 Flash via OpenRouter) to extract a navigable document tree with summaries, cross-references, and lazy-loadable content.

Production: https://aiapi.sciron.tech

Architecture

PDF Upload → Docling Sidecar (Python, :3001) → Rust API (:3002) → OpenRouter/Gemini → Structured JSON
                                                      ↓ (optional)
                                                   Supabase

• Docling Sidecar — Python FastAPI service for PDF-to-text (OCR + markdown). Lazy-loads ML models on first request.
• Rust API — Axum server that orchestrates the extraction pipeline and serves results.

Setup

make setup
# Edit .env and set OPENROUTER_API_KEY (required)
# Optionally set SUPABASE_URL and SUPABASE_SERVICE_ROLE_KEY for persistence
# Optionally set PORT to change the API port (default: 3002)

Run

make run

This starts:

• Docling sidecar on http://localhost:3001
• Rust API on http://localhost:3002 (configurable via PORT env var)

API

Endpoint	Method	Description
/health	GET	Health check
/configs	GET	List available extraction configs
/configs/:name	GET	Get a specific config
/extract?config=legal_br&upload=true	POST	Upload PDF (multipart file field), run extraction. upload=true persists to Supabase.
/extractions	GET	List all extractions (lightweight summaries with IDs)
/extractions/:id/snapshot	GET	Full extraction tree in one call (no raw content blobs, optimized for MCP/context loading)
/extractions/:id	GET	Get extraction by ID
/extractions/:id/node/:node_id	GET	Get specific node
/content/:ref	GET	Lazy-load content (supports ?offset=0&limit=4000)

Example

curl -X POST https://aiapi.sciron.tech/extract \
  -F "file=@document.pdf" \
  -G -d "config=legal_br&upload=true"

Configs

Domain-specific extraction configs live in configs/*.json. Each config defines:

• LLM prompt for structure extraction
• Allowed node types and subtypes
• Relationship types
• Metadata schema

Currently available: legal_br (Brazilian legal case files).