markusstrasser corpus — canonical local paper store
The content-addressed paper store lives at ~/Projects/corpus/. It holds source
bytes, immutable parser-addressed text, metadata, and parse-health state. Check it
before fetching or parsing a paper again.
The implementation and CLI belong to
~/Projects/research-mcp/src/research_mcp/corpus/. Substrate is dormant and is
not a runtime dependency.
CLI
research-mcp is installed as an editable uv tool and exposes corpus on
PATH. Do not use uvx corpus; that name belongs to an unrelated PyPI
package. The root is always explicit:
corpus --corpus-root ~/Projects/corpus stats
corpus --corpus-root ~/Projects/corpus status
corpus --corpus-root ~/Projects/corpus lookup --doi 10.1000/example
corpus --corpus-root ~/Projects/corpus lookup --pmid 12345678
corpus --corpus-root ~/Projects/corpus show doi_10_1000_example
lookup is the cheap presence probe. show returns metadata and the active
parsed path. Read parsed.<parser_id>/page.md from that returned path.
Ingest
corpus --corpus-root ~/Projects/corpus ingest --pdf paper.pdf --doi 10.1000/example
corpus --corpus-root ~/Projects/corpus ingest --url "https://pmc.ncbi.nlm.nih.gov/articles/PMC123/" --pmid 12345678
corpus --corpus-root ~/Projects/corpus ingest --reparse --paper-id doi_10_1000_example --parser marker-modal
corpus --corpus-root ~/Projects/corpus ingest-batch --dir papers/ --parser marker-modal --max-parallel 5
For papers and preprints, marker-modal is the default: GPU Marker plus Gemini
cleanup. Use pymupdf4llm for a cheap local text-first parse, trafilatura
for HTML, or mineru only when layout recovery justifies it. Existing parses
are immutable and parser-addressed; reparse adds a new parse and repoints the
active one.
Identity priority is DOI, then PMID, then content hash. Never invent identifiers or rename a source directory by hand.
Research MCP
The same owner exposes paper lookup/fetch/read tools through research-mcp. Use
those when an MCP result is more convenient; use the CLI for filesystem-oriented
inspection and explicit ingest. Both paths write the same ~/Projects/corpus
store.
Boundaries
- Corpus annotations, cross-repo attestation, outboxes, claim relations, graph maintenance, and sync jobs were retired on 2026-07-14.
- The corpus is a local paper cache, not a verdict store, remote service, or backup system.
- Never edit source bytes, metadata, or parsed outputs in place. Ingest a revision or reparse through the CLI.
- Web discovery belongs to
/research; this skill starts once a DOI, PMID, source ID, URL, or local source file is known.
Related skills
research— discover papers, then promote useful sources into this store.bio-verify— verify biological constants against authoritative databases; corpus can retain the supporting paper.