ADR-00001: PDF parsing backend — liteparse over docling¶
Context¶
parse_full_text needs a ParserBackend that converts fetched PDF full text (arXiv, and the local filesystem source) into Markdown. Several viable PDF-to-Markdown libraries exist, differing mainly in fidelity of structure recovery versus install weight — see Non-functional requirements → Dependency selection for the general criteria this and every other backend choice is held to.
Decision¶
LiteParsePdfBackend, backed by liteparse.
Alternatives considered¶
docling(and comparable heavier ML-based document-parsing libraries) — rejected specifically because it pulls in a PyTorch dependency, a materially heavier install footprint than a project that has otherwise stayed light on dependencies wants to carry for its default backend.
Consequences¶
v1's PDF path has no PyTorch dependency. Higher-fidelity structure recovery on complex layouts (tables, multi-column academic formatting) is a known, accepted ceiling of liteparse relative to docling-class tools — not solved here, but left as a named future option: SRS overview → Out of scope for v1 already names a slower, more sophisticated PDF backend as a future addition, swappable in via the same pluggable ParserBackend interface without changing the interface itself.