Parse configurations are foundational for quality of your outputs, and we see some common scenarios in which certain configurations would be helpful.

Is there handwriting or small text? Use Agentic OCR mode.

Agentic OCR is much better at parsing handwriting correctly as it can correct small typos or mistakes. Agentic mode is a good fit when a tradeoff of latency/cost for improved accuracy is reasonable.

Are there non-Germanic languages? Use the Multilingual OCR system.

Use the Multilingual or Hybrid options when parsing languages other than English, Spanish, Italian, Portuguese, French, or German. It also works well for picking up unicode characters.

Are there unexpected symbols in the output not visually present on your document? Check your Extraction mode.

Metadata embeddings are oftentimes corrupt for PDFs or contain unexpected information. Using Hybrid or Metadata Extraction modes can throw off your parse output since they use the document’s metadata. Stick to OCR mode if you’re not sure if there are good quality metadata embeddings.

Want checkboxes or images in the output? Check out our experimental options.

We have experimental configurations that let you enable_checkboxes, return_figure_images, return_table_images, rotate_pages, enable_scripts, and much more.

Complicated tables or table structure? Play with enrich’s table mode, or ai_json.

If your row or column structure is misaligned, try enabling enrich with table mode. Use the prompt to add details on what should be fixed. The ai_json option for Table Output Format passes an image of your table to a model for analysis, increasing latency but sometimes providing better outputs.