What it does
Docpick is a schema-driven extraction pipeline that combines local OCR engines with local LLMs to turn a document into structured JSON. It ships with eight built-in schemas covering commercial invoices, receipts, bills of lading, purchase orders, Korean e-tax invoices, bank statements, passport and ID documents, and certificates of origin. Custom schemas can be defined as Pydantic models or as JSON Schema files. Validation includes check-digit algorithms such as Luhn, ISO 6346, IBAN mod 97 and MRZ, cross-field rules like sum and date-order checks, and consistency checks across related documents.
The tool suits cases where fields from scanned or digital documents need to be pulled into a fixed structure and checked against the document's own internal logic, for example invoice line-item totals or container weight sums. It also handles Korean text through EasyOCR and can process whole directories with parallel workers. When an OCR engine fails, the pipeline falls back to the next available engine. If LLM output cannot be parsed, it retries once with a correction prompt, and partial results are returned with errors recorded in the result.
Running it requires Python 3.11 or later and an LLM endpoint such as vLLM, Ollama or another OpenAI-compatible service. Providers and models are set through the CLI or a YAML file. The default auto OCR engine runs PaddleOCR or EasyOCR on the CPU first and escalates to GOT-OCR2.0 or a vision-language model on the GPU when average confidence falls below a threshold, which defaults to 0.7.
Limitations and requirements
Requires Python 3.11 or later and a running LLM endpoint (vLLM, Ollama or OpenAI-compatible). GOT-OCR2.0 and the VLM engine require a GPU; PaddleOCR and EasyOCR can use a GPU optionally. The auto engine's escalation to GPU-based engines depends on the confidence threshold being met.
Safety
Our team has not reviewed this resource. Only automated rules checked it; resources with high-risk signals are not published. These are signals, not a guarantee that installing it is safe. Try it in a sandbox project first.
- Runs code or changes an assistant's behaviour
Sources
- Official MCP Registry first seen 13 Nov 2026 · last seen 13 Nov 2026 · its page on the source
Docpick
Schema-driven document extraction with local OCR + LLM. Document in, Structured JSON out.
v0.1.2As shown on registry.modelcontextprotocol.io when we last read it