musharna/data-aggregator-mcp

Destination dir (default managed cache)

Unpack downloaded zip/tar archives into the destination (default false). Path-traversal-guarded; counts against max_bytes.

Glob over file names (default all)

Override max_bytes

Source-prefixed id or bare Zenodo id

Byte ceiling before failing loud

When true, probe 5 sources (zenodo, datacite, omics, literature, huggingface) and attach a 'health' field ({status: up|down, latency_ms, detail}) to those entries; every other source gets health: null. Default false: returns the static catalog with no network.

Optional column projection for head.

File name within the record; optional when exactly one operable file is present.

DataResource id (e.g. 'zenodo:7654321')

Row count for head/preview

Read-only SELECT for op='sql'.

2-10 source-prefixed resource ids to relate.

Optional citation format to render onto the result: 'bibtex', 'ris', 'csl-json', or any CSL style name ('apa', 'mla', 'vancouver', ...). DOI-bearing records render via DOI content negotiation; non-DOI records support 'csl-json' only. Omitted = no citation. Failures degrade quietly (citation stays null).

When true, attach an RDA-grounded FAIRness assessment under fair{}: a 0–100 overall score plus findable/accessible/interoperable/reusable sub-scores, the count of indicators evaluated, and actionable gaps each naming its RDA FAIR Data Maturity Model indicator id. Pure/local — no network call. Only the machine-evaluable subset is scored (never fabricates what the metadata cannot show).

Optional export to render onto the result. 'croissant' attaches a file-level Croissant JSON-LD manifest (croissant field); 'ro-crate' attaches a minimal RO-Crate 1.1 manifest (ro_crate field); 'provenance' attaches a one-call RO-Crate 1.1 data-availability dossier (provenance field) bundling version-currency, licence+SPDX, FAIR score, retraction status, and the source/DOI/ID chain — it auto-attaches fair{} and trust{} so the dossier is complete in one call (unknown signals are reported as unknown, never as a clean claim).

Source-prefixed id, bare Zenodo id, or DOI

When true, attach trust signals (retraction status via Crossref) to the result under trust{}. One extra Crossref call; only meaningful for DOI-bearing records (a DataCite data DOI Crossref does not register leaves retracted=null = unknown, never a false clean claim).

When set, attach a licence-compatibility advisory under license_compat{} for an intended use of the record. Supported intents: 'commercial', 'redistribute', 'modify', 'ml-training' (training = a derivative+commercial use, our stated interpretation). The verdict is ALLOW/REVIEW/DENY computed from a bundled choosealicense.com licence matrix keyed on the normalized SPDX id, naming the governing clause — a metadata-derived advisory, NOT legal advice. An unrecognized or absent licence yields REVIEW (never a fabricated ALLOW/DENY); an unknown intent is an error.

Optional assay/method name. Resolved via EDAM topics (EBI OLS); the query is expanded with the canonical name + exact synonyms (e.g. 'ChIP-seq' also matches 'ChIP-sequencing'/'ChIP-exo'). An unknown term yields no expansion; an OLS failure surfaces in errors. The expansion is echoed in assay_expansion.

Optional chemical/compound name. Resolved via ChEBI (EBI OLS); the query is expanded with the canonical name + exact synonyms (e.g. 'caffeine' also matches '1,3,7-trimethylxanthine'), capped to a bounded number of synonyms. An unknown term yields no expansion; an OLS failure surfaces in errors. The expansion is echoed in chemical_expansion.

Opt into conservative cross-repo content dedup (default false). On top of the always-on exact-DOI dedup, folds records that are the SAME dataset deposited under different (or no) DOIs — e.g. a Zenodo mirror of a figshare deposit, GEO<->ArrayExpress — into one record, annotating the survivor with the folded copies under mirrors[]. Conservative: a merge needs a shared file checksum OR identical (normalized-title, first-author-surname, year); title-only or partial matches never merge. Intra-page / best-effort only (a mirror on a different page is not collapsed), so a page may return fewer than size items; pagination is unaffected.

Opaque pagination token from a prior search's next_cursor. When set, all other search params are read from the cursor.

Optional disease/phenotype name. Resolved via MeSH (NCBI E-utilities); the query is expanded with the canonical descriptor + entry-term synonyms (e.g. 'breast cancer' also matches 'Breast Neoplasms'). The expansion is echoed in mesh_expansion.

Keep only results of this kind.

Opt into diverse multi-query recall expansion: an LLM generates up to a few deliberately-diverse reformulations of your query, each is fanned out across all sources, and the deduped union is re-ranked against your original query — surfacing relevant records a single keyword query would miss. Costs N× the upstream calls (bounded). Requires an LLM endpoint (LLM_API_BASE); with none configured the search runs as a normal single query and notes it in errors['multi_query']. The variants used are echoed in query_expansion. Composes with understand=. NOTE: multi_query=true ALWAYS applies semantic re-ranking of the window internally regardless of rank=; the rank= param has no effect in this mode.

Optional organism name. Resolved via NCBI Taxonomy; the query is expanded with the canonical name + synonyms (e.g. 'Orobanche aegyptiaca' also matches 'Phelipanche aegyptiaca'). The expansion is echoed in taxon_expansion.

Opt into a whole-search RO-Crate 1.1 Run Crate (default false). Attaches provenance_crate{} — a machine-readable manifest documenting this search: the query, the sources queried, the ontology expansions that fired, the per-source errors (a partial search is disclosed), and per-hit provenance for every result (version-currency, licence + normalized SPDX, FAIR score). Per-hit RETRACTION is omitted — it would need one Crossref call per hit; use per-record resolve(format=provenance) for that. Covers THIS search page only (intra-page; each page of a paginated search gets its own crate).

Keep results with year >= this.

Keep results with year <= this.

Free-text search query

Result ordering. 'relevance' (default) = upstream/merged order. 'semantic' re-ranks the fetched page by embedding similarity to the query (needs EMBEDDING_API_BASE; degrades to relevance order with an errors['semantic'] note if unconfigured). In semantic mode pagination is window-based (each page consumes its full fetched window).

Max results (1-50, default 10)

Restrict fan-out to these sources (default: all). Available: zenodo, dataone, gbif, cellxgene, datacite, dandi, omics, literature, huggingface, datagov, nasacmr, omicsdi, openml, pdb, uniprot, gwas, biostudies

Optional tissue/anatomy name. Resolved via UBERON (EBI OLS); the query is expanded with the canonical term + exact synonyms (e.g. 'liver' also matches 'iecur'/'jecur'). The expansion is echoed in tissue_expansion.

Opt into LLM query understanding: a free-text query is rewritten into a keyword core + structured params (organism/disease/tissue/chemical/assay, kind, year) before fan-out; extracted entities are validated by the same ontology resolvers (a hallucinated entity that doesn't resolve is simply dropped), explicit params you pass always win, and the interpretation is echoed in query_understanding. Requires an LLM endpoint (LLM_API_BASE); with none configured the search runs unchanged and notes it in errors['understand'].