docStar – Configurable Multi-Domain Document Checking System

THE IDEA

docStar – Configurable Multi-Domain Document Checking System

Type: Software Status: Beta Tech Stack: FastAPI, Python 3.12, PostgreSQL 16 (RLS, JSONB, tsvector), SQLAlchemy, Alembic, Celery, Redis, IBM Docling + EasyOCR, Ollama/vLLM/OpenAI, React 18, Vite, TypeScript, TanStack Query, Tailwind, Recharts, Nginx, Docker Compose, Prometheus/Grafana/Alertmanager, Keycloak/OIDC, Fernet (PII encryption) Problem Statement A document inbox can only be evaluated automatically when three things come together: clearly defined fields per document type, a control instance for uncertain results, and an answer layer that can be traced back. Common solutions presuppose a cloud service for this, are tailored to a single industry and let AI answers stand without evidence — for documents containing personal data that is not a workable basis. Description docStar receives incoming documents, captures them via an OCR microservice (IBM Docling + EasyOCR) and evaluates them schema-driven with a provider-agnostic LLM (Ollama locally, vLLM or OpenAI): a category registry defines, per document category (twelve active categories from invoice to reference letter, plus a legacy category), exactly which fields are expected, and derives from that a strict Pydantic validation model against which the AI output is checked. Low confidence, unconfirmed fields or detected prompt injection (two-stage via regex + BERT) automatically route a document into a human-in-the-loop review queue. The reviewed data is full-text searchable, can be aggregated as office insights, exported (CSV/JSON/ZIP) and queried through a natural-language chat that, where useful, returns structured answer tables with versioned citations and stores conversations as sessions — without free SQL, through three fixed, type-safe tools. The system is domain-independent and configurable via config/domain.json (active: corporate), multi-tenant capable through Postgres row-level security with privilege-separated database roles, and stores personal data encrypted at rest (Fernet/MultiFernet with key rotation) with built-in GDPR lookup and erasure. Its distinguishing traits are fully local operation without any cloud data outflow, the domain-independent schema-driven extraction, the citation-backed chat answer layer and an unusually openly documented, ongoing security and correctness audit discipline. Use Case Companies automate their document inbox (invoices, contracts, HR records and more) — capturing, structuring, checking, evaluating and querying it in natural language with evidence — data-sovereign and without being forced into the cloud. Link: https://github.com/rawk7000/DocStar (private repo)