MARE turns PDF QA into evidence payloads

The abstraction is useful because most PDF RAG systems stop at chunks and citations, while MARE makes the retrieval artifact itself the product. The sharp edge is that trust only holds if table, figure, and abstention behavior get as strong as its page-and-snippet path.

• –The repo is opinionated in the right way: it explicitly says it is not trying to be a full agent framework, vector DB, or parser platform, which makes the evidence-layer positioning clearer than many bloated RAG stacks.
• –The current payload shape already fits agent workflows well because page image, localized highlight, and rationale are easier for downstream systems to verify, rank, or show to humans than plain text chunks with a citation string.
• –The most valuable next step for agent compatibility is stronger structured provenance: stable page/object IDs, bounding boxes, confidence or abstain scores, and machine-readable evidence spans that can survive retries and multi-step tool use.
• –The repo is candid that table retrieval, figure retrieval, and deeper layout-aware routing are still early, and those are exactly where many PDF systems break once documents stop looking like clean prose.
• –An evaluation harness for page, object, and no-result benchmarking is a strong sign of seriousness, because abstention and evidence quality usually matter more than answer fluency in real PDF workflows.