Shared Evidence LibraryCopy Anchor Link

A submission form plus a free-text/kind search panel over the team-wide evidence repository — cut cards and reusable analytic blocks.

• Route: /research/cards/library
• Package: debate-card-search

What it showsCopy Anchor Link

A submission form (kind, topic, case area, argument block, citation, tags, body text with a live word count) plus a search box, kind filter, and topic/case-area/tags filters over every persisted EvidenceLibraryEntry. Each result has Edit and Delete actions; editing records a scored CardRevisionRecord for the Revision Incentives leaderboard. A "Check this page" box checks whether a pasted URL has already been cut, both locally and against a shared, server-backed reuse index consumed by a browser extension (apps/debate-web-ext).

The Argument Library browser (/research/cards/argument-library) composes this repository with tagged Contributions Feed submissions into topic/case-area folders and cross-cutting tag collections, with rename/merge and bulk tag-editing tools.

Real search indexCopy Anchor Link

A real inverted-index, TF-IDF-ranked search (lib/evidence-search-index.ts) replaced the original keyword-overlap scan, cached across calls and updated incrementally rather than rebuilt on every write.

Opening a pre-filled searchCopy Anchor Link

/research/cards reads its starting search from the URL: ?q= (term), ?year=, ?event= (CX, PF, LD, NDT, NFA) and ?school= / ?team= / ?tournament=. Other pages build these links with buildCardsSearchHref() and useSearchState reads them once on mount via readCardsSearchParams() (both in lib/search-query.ts). The topics explorer on /practice/statistics uses this to link every resolution to its cards. The URL isn't kept in sync as you edit the search afterwards.

The first search is ready on openCopy Anchor Link

Opening /research/cards with no pre-filled search shows the default list (newest cards, quotes) from your last visit straight away, kept in localStorage by lib/default-search-cache.ts, while the page fetches the current list and swaps it in without a spinner. That first request skips the 300 ms typing debounce. On the server, /api/search keeps the no-term, no-filter list in each Worker isolate for five minutes, since it sorts the debate_cards table by imported_at. Every request still counts toward the daily search limit.

A new search always winsCopy Anchor Link

Typing waits out the 300 ms debounce, but changing the term, sort or filters retires the previous search at once in useSearchState: its request is aborted (plain fetch with an AbortController, since grab keys cancellation by path and every query is a different path) and its answer is dropped by a request sequence number, so results for an older query can't land after a newer one. A committed query — Enter, a clicked suggestion, or an accepted inline completion — calls setSearchTerm(term, { immediate: true }) and searches with no debounce; if that term is already waiting on the debounce, the pending search goes out now.

The search box (ui/primitives/autocomplete.tsx) shows the top completion inline as ghost text (inlineCompletionSuffix() in lib/search-query.ts), accepted with Tab or →. ↑/↓ load the next suggestion into the box and its results follow after the debounce, so stepping quickly through the list costs one search, not one per suggestion. Suggestion requests carry their own sequence number, so a slow answer for an earlier prefix can't replace the list for the newer one. Results are not prefetched for suggestions the user hasn't reached, because every request counts toward the daily search limit.

Data flowCopy Anchor Link

panels/EvidenceLibraryPanel.tsx
  → saveEvidenceLibraryEntry() / saveEvidenceLibraryEntryRevision()
  → searchPersistedEvidenceLibraryWithIndex(query) — state/evidenceLibraryEntries.ts
  → checkPersistedPageForExistingCards(url)

Known gapsCopy Anchor Link

The imported_at index that turns the default list into an index walk is created by the weekly cron or an admin search-index action (see card library import); until one runs, the uncached default list is still a full-table sort.

Tag identity is still exact-string, though autocomplete matching is case-insensitive, typed tags are normalized to an existing casing, and the library browser surfaces/merges any casing variants that slip through.