Local AI RecipesOCR / documents

PP-OCRv6-small local document OCR recipe

A production-oriented OCR baseline for private local documents, where real scans, layout recovery and regression matter more than leaderboard novelty.

AILocalabs production baselineReviewed 2026-09-26
Commercial statusPaddleOCR is an Apache-2.0 project; continue checking exact packaged models and third-party dependencies whenever the bundle changes.
HardwareCPU/OpenVINO and GPU paths should be benchmarked on real documents. Small models are attractive when OCR is not allowed to dominate the full translation pipeline.
RuntimePP-OCRv6-small is the current LexKiln OCR production baseline.

For document products, OCR accuracy is only one layer. The useful benchmark includes reading order, tables, formulas, page regions, downstream translation and final layout recovery.

AILocalabs therefore treats new OCR releases as candidates only when they beat the production baseline on actual documents.

DIY recipe

Build it yourself

  1. Build a representative set of real scans: medical PDFs, invoices, mixed-language pages, tables and low-quality images.
  2. Measure OCR accuracy together with CPU latency, memory and downstream layout quality.
  3. Keep a golden regression set before changing OCR model/version.
  4. Prefer incremental replacement of weak stages instead of rewriting the entire document pipeline for every new model.
Risks

What can go wrong

  • OCR benchmark wins can disappear on low-quality scans or unusual typography.
  • A faster OCR model can still reduce total product quality if reading order/layout metadata gets worse.
  • Medical and private documents need a clear local-data handling story in addition to model quality.
For humans and AI agents

Give the recipe to your AI and reproduce it locally

Target hardware, runtime, key parameters, validation and risks are published as Markdown and JSON. An AI agent should still inspect your machine first, explain changes and ask before risky operations.

AILocalabs decision

What we ship — and what we do not

LexKiln already uses PP-OCRv6-small as its production OCR baseline. Recipe updates should come from real failure cases and measured improvements.

Upstream / license sourcesPaddleOCR repository
Recipe discussion

Tried it? Leave results, questions and improvements.

Discussion stays with this recipe so future builders can learn from it. Never post passwords, API keys, license keys or private files.

Loading discussion…

Join the discussion