Dual-VLM OCR
Every region is OCR’d by PaddleOCR-VL-1.6 and GLM-OCR in parallel. Paddle priority on agreement; GLM as fallback and cross-check; per-block risk flags surface advisory data.
docstack-ocr
Dual-VLM OCR + tenant-authored extraction. Self-hosted, model-agnostic, audit-first.
Built for system integrators
Section titled “Built for system integrators”docstack-ocr is a self-hosted document understanding platform designed to plug into a larger system — not to be the system. Submit a document, poll for completion, retrieve a rich CanonicalOcrDocument plus structured fields conforming to a tenant-authored template.
Tenant-authored templates
Document types are runtime-defined. JSON Schema (Draft 2020-12) plus a declarative validation DSL (required, regex, arithmetic, date_plausible, enum, checksum, cross_agreement). Zero hardcoded templates.
Two auth modes
bearer (built-in users + API keys + sessions + CSRF) or trusted_headers
(your gateway forwards X-Tenant-Id / X-Actor-Id / X-Actor-Roles).
Model-agnostic LLM
vLLM, Ollama, OpenAI, or native Gemini. Swap via runtime infrastructure config; no code changes.
Review-first
Validation failures route to a queue with approve / reprocess / field-level override. Every override re-runs validation and logs an audit row.
Auditable + recoverable
Every mutation is audited. Soft-delete is the default; restore is one idempotent endpoint away.
Where to go next
Section titled “Where to go next”- Quickstart — boot the stack, mint a key, submit your first document. ~5 minutes.
- Infrastructure setup — configure OCR and LLM endpoints across vLLM, Ollama, OpenAI, and Gemini.
- Full integration guide — every endpoint, every parameter, every failure mode. The single canonical markdown reference.
- API reference — interactive OpenAPI browser with a “Try it” panel.
- Contract testing — keep your integration honest by fuzz-testing the API against its own OpenAPI spec.
For AI agents
Section titled “For AI agents”Every page exposes a Copy as Markdown action and an Open in ChatGPT/Claude dropdown. The full docs surface is also exported in machine-friendly form:
/llms.txt— index following the llmstxt.org convention./llms-full.txt— the entire docs site flattened into one Markdown file./llms-small.txt— same asllms-full.txtbut trimmed for smaller context windows./integration-guide.md— the canonical single-file integration reference, raw./openapi.json— OpenAPI 3.1 spec snapshot.