Abbreviation search¶

Desktop search combines Nucleo abbreviation matching with OSA typo correction for Preferences and Move, plus phrase priority for indexed mail search. Sam selected this combination on 12 September after comparing all three demos.

Use a disposable profile to explore the fictional search examples:

shep_demo_dir=$(mktemp -d "${TMPDIR:-/tmp}/shep-search-demo.XXXXXX")
CARGO_BUILD_JOBS=4 cargo run --profile test-ui --features test-support -- --demo --search-mail --persist-demo --test-state "$shep_demo_dir/state.json"

Use a fresh directory for a new trial. Reusing the directory retains changes to the fictional workspace and allows the Profiles controls to open.

In Preferences, try prf, ntfctns, dark mode and appearnce. In Move, compare Archive, archvie and pjarch against folders named Archive and Projects/Archive. Mail needs cached messages to compare: a short body equal to the query comes first; a matching phrase then precedes separated exact words, which precede prefix or corrected words. For example, search architecture plan 17 among messages containing that phrase, 17 plan architecture and architecture planning 17. The number 170 does not match 17.

Exact labels, exact folder leaves and prefixes receive the strongest label scores. Every remaining query term must match. Nucleo rewards compact subsequences and word boundaries; if a term has no subsequence match, OSA allows one edit for words of at least four characters in Settings or three in Move, and two for at least six. Transpositions count as one edit. Label terms containing digits require a complete token, so S3 cannot match S30. Folder punctuation is passed literally to the low-level matcher; Preferences and mail split text into words. Neither enables !, ^, $ or | as query operators. One matcher and its scratch buffer are reused through each result batch. Preferences caches each query word's characters alongside the immutable catalogue words and their field weights. Preferences abbreviations match individual indexed words; folder abbreviations can span path segments, as in pjarch for Projects/Archive.

When Move also searches other accounts' folders, each foreign row adds 15 to its label score. That is more than the exact-leaf step of 10, so a folder in the current account whose leaf equals the query still comes first, and less than the prefix band starting at 20, so an exact or prefix match in another account precedes any substring, abbreviation or typo match at home. With arch against a local Archive and Archives, a foreign Archive and a local Projects/Archive, the order is local Archive (23), Archives (24), foreign Archive (38), Projects/Archive (48); with archive it is local Archive (0), Projects/Archive (10), foreign Archive (15), Archives (21). Foreign rows only appear while the search box has text.

Mail uses SQLite FTS5 and field-weighted BM25 within each tier. Subject, body and sender weights are 2.0, 1.0 and 0.3. Phrase and literal relations are computed once, with a shared relation for single-term queries. The list and frozen bulk selection use identical tier, score, timestamp and identity ordering; pages remain limited to 50 metadata rows.

Search totals and unread counts are computed alongside the ranked keys inside SQLite, then at most 50 metadata rows are read. An empty page still retrieves the complete counts. Queries containing only proven exact terms share their literal ranking instead of computing it again; phrase priority remains intact. Unicode exact-body matches retain their priority even when SQLite's tokeniser and the query normaliser differ for combined accents. For ASCII alphabetic words, the current vocabulary is checked inside the query transaction. If no longer term exists, an exact FTS lookup avoids unnecessary prefix merging. The proof is repeated for each query, so new word extensions remain searchable after insertion, removal and reopening the cache. When a bounded prefix scan exhausts its range, impossible spelling variants in that range are skipped. Other prefixes and scans that reach their cap retain their complete candidate checks.

Mail abbreviations do not scan every message or vocabulary term. Mail keeps the bounded expansion: the first 12 query tokens, 32 one-edit vocabulary candidates, 256 prefix-neighbour candidates for longer words and 12 corrections per token. Both candidate budgets count typo neighbours and exclude the exact token, which already has its own literal/prefix alternative. This can admit one more neighbour at the earlier cap boundary and avoids recounting the exact word's postings during vocabulary lookup. Further query tokens are ignored. Only ASCII alphabetic words receive typo expansion; Unicode literal matching remains available. Wholly numeric mail tokens are exact; mixed letters and digits retain prefix matching, so mail S3 can still match S30. Latin accents are normalised without stripping Japanese marks or decomposing Hangul. BM25 statistics use the cached collection, while account, folder and flag filters still restrict returned results.

Algorithm references: Nucleo matcher, RapidFuzz OSA and SQLite FTS5.

Automated ranking and capture coverage lives in tests/search_abbreviations.rs, alongside the shared search and selection suites. Native demonstrations and performance receipts are recorded in the completion log after review. On the shared 100,000-message fixture, cached search p95 is 19.87 ms for ordinary text, 22.83 ms for a transposition and 31.25 ms for four terms. Warm complete Preferences ranking is 0.023–0.045 ms p95; these measurements exclude native input/drawing and live providers. Browser and mobile do not acquire this experimental matcher by switching desktop branches; their search parity remains tracked separately.