Jina metadata parser is blog-only

Why Care?

The operator has been hand-entering metadata that Jina already returned — “supremely inefficient.” A live fetch of a Springer article proved it: Jina returned authors (dc.creator, citation_author), publisher (citation_publisher = Springer US), and date (dc.date = 2023-05-27, prism.publicationDate, citation_publication_date) — but the parser only looked for blog/OpenGraph keys (article:published_time, meta.date, og:site_name), which academic pages don’t set. So date + publisher fell through empty. A Jina fetch of a company landing page is different again (og/site metadata, no author or date at all).

The design (operator-directed)

Not one parser reaching for more keys — two parser profiles plus fuzzy routing, so when auto-detection is wrong the human can press a button to re-parse under the other profile:

• structured — Highwire citation_*, Dublin Core dc.*, PRISM prism.*. For scholarly articles, journals, anything with a DOI.
• opengraph — og:*, article:*, Jina top-level, generic. For blogs, news, company landing pages, plain web pages.

Routing detects the source kind (academic-paper / article / company-landing / web-page) from which keys are present, records it on the source, and picks the profile. A forceProfile override lets the operator re-route (button wiring tracked as a fast-follow).

Each profile resolves every field across an ordered alias list (first hit wins), so a field missing under one convention still resolves under another. Different kinds expect different fields — a company landing page is not nagged for an author/date it doesn’t have.

Authors

Authors stay an array of strings (Jina’s citation_author / dc.creator already arrive as arrays; one array element = one author, never comma-split, so “Pal, Soumen” survives). The UI author field already coerces a comma-separated entry into an array. Future: array of strings that resolve to unique author profiles (the persons canonical layer).

Fix

jina.ts gains detectProfile() (routing → profile + kind), extractBib() (pure, profile-aware, forceProfile-overridable), and per-profile field alias lists. fetchViaJina() accepts forceProfile. Vitest fixtures cover the academic (real Springer metadata), article, and company-landing kinds.

Resolution

Fixed 2026-08-02. Verified: content-ingest vitest run + tsc --noEmit. Reaches augment.didi.sh on the next redeploy. The manual re-route button (UI + a re-extract capability that reuses the cached Jina result) is a fast-follow, tracked separately.

See also

• [[Fetch-Full-Content-Clobbers-Operator-Metadata]] — the sibling fetch fix.
• [[feedback_additive_enrichment_never_overrides_accepted]] · [[feedback_human_in_drivers_seat]]