Chunk output; custom prompts; structured extraction improvements

Marker 1.8.0

• Marker will now output a flat list of blocks with associated html, which is useful for RAG
• Structured extraction beta is significantly improved, with better performance/accuracy
• New LLM sectionheader processor will correctly label section header levels
• You can pass a prompt to marker in LLM mode to adjust the output
• Marker batch conversion script has somewhat better performance, closer to our inference container - email us at hi@datalab.to if you want to get setup with our inference container (used on prem at top AI research orgs)
• Add an option to filter out blank white page images from output
• Enable keeping pageheader/pagefooter.

Chunking/RAG improvements

• Add chunk output format which is a flat list of chunks with full html in each
• Add an llm sectionheader processor that will redo all the header levels against each other properly

Use the sectionheaderprocessor by setting --use_llm, and the chunk output by setting --output_format chunks.

Structured extraction

• Fix structured extraction, so it works much better than before (requires llm)
• Improve structured extraction test app

You can try with with the streamlit app by running python extraction_app.py.

Promptability/customization!

• Add promptability via block_correction_prompt, which can be used to create custom behavior (requires llm)

Try it by setting the block_correction_prompt config key to a specific prompt.

Misc

• Get the marker script to perform a bit closer to our inference container by default (inference container gets 10-25 pages/s on H100). Will auto-configure worker count to available VRAM.
• Fix where marker would output blank pages as images
• Enable keeping pageheader/pagefooter in the output
• Adjust llm services to enable text-only input
• Add html field to almost every block type

Test pageheader/pagefooter by setting keep_pagefooter_in_output and keep_pageheader_in_output.

What's Changed

• Adding image format arg to OpenAI interface by @rgeorgi in #752
• fix: don't remove leading newlines when paginate=True for markdown by @zanussbaum in #769
• Fix LLM query retry logic and unnecessary sleeping by @runarmod in #772
• Update issue templates by @tarun-menta in #767
• WIP: Add Blank Page Processor by @tarun-menta in #750
• Vik/0625fixes by @VikParuchuri in #768
• Marker v1.8 by @VikParuchuri in #773
• Add Azure OpenAI service support to marker package by @MauritsBrinkman in #675
• Add azure openai by @VikParuchuri in #774

New Contributors

• @rgeorgi made their first contribution in #752
• @zanussbaum made their first contribution in #769
• @runarmod made their first contribution in #772
• @MauritsBrinkman made their first contribution in #675

Full Changelog: v1.7.5...v1.8.0