// the find
NanoNets/docstrange
Extract and convert data from any document, images, pdfs, word doc, ppt or URL into multiple formats (Markdown, JSON, CSV, HTML) with intelligent structured data extraction and advanced OCR.
DocStrange is a document-to-structured-data converter from Nanonets that handles PDFs, DOCX, images, spreadsheets, and URLs, outputting Markdown, JSON, CSV, or HTML. It runs either against their cloud API (10k docs/month free) or locally on CPU/GPU using a 7B model. The primary audience is developers building RAG pipelines who need clean text out of messy scanned documents.
The dual-mode design is practical — cloud for quick prototyping, local GPU for privacy-sensitive workloads, and the fallback to CPU means it actually runs without special hardware. Schema-driven extraction (pass a JSON schema, get back conforming JSON) is genuinely useful for invoice/contract automation without training a custom model. The CLI is well-thought-out: field extraction, schema files, batch globs, and output-to-file all work as first-class options, not afterthoughts. The MCP server integration for Claude Desktop is a nice touch for document-heavy workflows, even if it's dev-only.
The cloud dependency is a real gotcha — the default mode silently sends your documents to Nanonets' servers, and the rate-limited anonymous tier is vague enough that you won't know you've hit it until something breaks. Local GPU mode requires downloading a 7B model, but there's no documentation on which model, its size, or licensing, which makes it hard to evaluate for compliance use cases. No async API: every `extract()` call is synchronous and blocking, which means you'll need to wrap it yourself for any concurrent workload. The test suite is almost entirely debug scripts and manual tests against real files — there are no unit tests for the processing pipeline, so you're taking the accuracy claims on faith.