Product · Document Intelligence Platform

The document platform that never phones home.

OCR, classification, field extraction, hybrid semantic search, and a grounded RAG copilot — engineered on-premises first for organizations whose documents can't leave the building, with pluggable cloud providers when they can.

Book a demo →See capabilities

quantpi · ai-dms/status

$ ai-dms.status()
✓ stack: 100% open components · models: local via vLLM/Ollama
✓ deployment: on-prem first · cloud-pluggable
✓ ip & data: stays inside your boundary
# productized from patterns we shipped repeatedly

PRD/01Capabilities

What it does

Ingestion & OCR

40+ formats through PaddleOCR and layout models — scans, photos, tables, and multi-language documents normalized into structured content.

formats: 40+ · langs: 100+

Classification & intelligence

Document-type classification, packet splitting, and metadata enrichment — the right pipeline processes the right pages automatically.

routing: automatic

Field extraction

Schema-driven extraction with per-field confidence and validation rules; low-confidence items route to human review, corrections feed evals.

confidence: per-field

Hybrid semantic search

BGE-M3 dense vectors fused with keyword search across the corpus — answers cite page and region, every time.

citations: page + region

RAG copilot

Ask your archive questions in natural language; grounded answers from your documents or an honest ‘not found’ — never an invention.

grounding: cited or refused

Security & governance

Keycloak SSO, role-based access carried into retrieval, PII detection via Presidio, malware scanning, and immutable audit logs.

access: ACL at query time

PRD/02Architecture

Inside the platform

Ingestion (ING)

NATS JetStream event backbone, format normalization, OCR orchestration, and MinIO object storage — built for sustained bulk throughput.

Processing (PRO)

Layout analysis (LayoutLMv3/Donut), table extraction, and enrichment pipelines with per-stage observability.

Classification & Intelligence (CLS)

Type classification, splitting, metadata models — eval-gated, retrainable on your corrections.

Search & Retrieval (SRH)

Qdrant/pgvector + OpenSearch hybrid retrieval with reranking, ACL filtering, and the RAG answer layer.

Platform layer

Keycloak, OpenBao secrets, Prometheus/Grafana/Loki/Tempo observability, MLflow model registry — operations-grade from install.

PRD/03Questions, answered straight

FAQ

Does AI-DMS require any cloud services?

No. The reference deployment is fully air-gapped: local models via vLLM or Ollama, BGE-M3 embeddings, Qdrant or pgvector, OpenSearch, MinIO — nothing leaves your network. Cloud model providers (Azure, AWS, GCP) are pluggable through the abstraction layer when policy allows.

What hardware does it need?

A meaningful pilot runs on a single GPU server (e.g., one A10/L40-class card) plus standard compute for the pipeline. Production sizing depends on document volume and model choice — we provide a sizing calculator during evaluation.

How accurate is extraction on our documents?

We answer that with your documents, not a brochure: evaluation includes a measured pilot on a sample corpus with per-field accuracy reported against a golden set you approve.

Can it integrate with our existing DMS or ERP?

Yes — API-first design with webhook, queue, and batch connectors. Common patterns include feeding extracted fields into ERPs and serving search over content that stays in your existing repository.