New features, improvements, and announcements.
The parser now emits two formats from a single parse pass: bgraph.json for machines and bgraph.md for humans and LLMs. Both carry identical content. Strip the metadata fences from bgraph.md and you get clean, readable Markdown any LLM can consume directly — no preprocessing needed.
The parser now accepts Markdown files alongside PDFs. Semantic structure, bounding-box locations where applicable, and full graph output — same pipeline, new input channel.
Section detection recall went from 61.6% to 98.6% with precision held at 1.000 on our benchmark corpus. Overall corpus micro F1 moved from 0.511 to 0.710. The parser is now meaningfully more reliable on documents with complex layout, numbered sections, and mixed bold weights.
Submit your Bragi project and get featured. We want to see what you're building.
Bragi is live! Try the parser, explore the API, and start building GraphRAG applications.
Want updates? Star us on GitHub or check back here.