Filler words, bookmark titles, and look-alike rows — three quiet refinements that make the first screen of results more useful.

Good search is a balance between two forces.

Recall brings candidate pages in: anything that might be what you mean. Precision puts them in order, so the page you actually want sits at the top of a short list. Most search improvements push on one side, and the best ones arrive in pairs.

SpotRecall, my private on-device browser recall extension, has spent several releases building recall. It matches any query word rather than requiring all of them, so long Japanese and Chinese phrases still find pages. It expands prefixes, so gar finds gardening. And since 1.1.0 it seeds the index with your bookmarks and open tabs, so useful pages are searchable from the first minute, before they have been revisited and captured.

Version 1.3.1 adds the precision side of each of those choices. None of the changes is large, and the first one below was the largest measured step in the release.

1. Bookmarks that match the whole query rank as they deserve

When SpotRecall imports a bookmark, at first it knows only the title and URL. The page content arrives later, the next time you visit. Until then, the record is title-only.

Titles are short, and keyword search rewards a term that makes up a large share of a short field. SpotRecall's BM25 index also weights the title three times higher than body text, which is the right choice for captured pages. Together, this means a title-only record that shares a single word with the query can score surprisingly well.

One example from the test suite: the query 门牌 网购 (doorplates, online shopping) against a bookmark titled 买手网,您身边的网购帮手 (a shopping-assistant site). The bookmark shares 网购, but it is not about doorplates at all.

1.3.1 adds a query-dependent prior for title-only records:

title-only record: title + URL contain every term of some query variant → weight 1.0 otherwise → weight 0.5 pages with captured content → always 1.0 

The phrase "of some query variant" matters. Each query is also searched in translation, so a bookmark that fully matches the English or Japanese version of the query keeps its full weight. A genuine title match — Japan Patent Office for the query japan patent — is untouched. Only the partial overlaps step back, letting pages that actually contain your words come first.

This was the largest measured step in the release. On 54 real queries against a 1,088-record backup, MRR rose from 0.809 to 0.828. In a worst-case simulation, where every backfilled bookmark also receives SpotRecall's bookmark boost as it would in a bookmark-heavy profile, MRR rose from 0.787 to 0.826.

The three behaviours are pinned by tests: a stray single-term match is down-weighted, a full match on any variant keeps full weight, and records with captured content are never touched.

2. Focusing on the words that carry meaning

Matching any query word is what makes SpotRecall forgiving with natural-language and CJK queries. It also means every word in the query brings its own candidates.

Most words bring useful ones. A few bring almost everything. The Chinese particle 的, the Japanese particle の, and English words like of and the appear in a large share of all pages. With prefix matching on top, of would even expand into office and offer.

1.3.1 handles this on the query side, with two small rules:

  • A short stopword list for English, Chinese, and Japanese (the, of, to…; 的, 了, 是…; の, は, を…) is removed from the query before matching — unless the query consists of nothing but stopwords, in which case it is searched as typed.
  • Prefix expansion applies to specific terms only: at least three characters for Latin script, at least two for CJK. gar still finds gardening, but of stays of.

From the test suite:

queryTerms('Gardening的日语翻译') → ['Gardening', '日语', '翻译'] 'Japanese translation of Gardening' → does not match "Office 365" queryTerms('the') → ['the'] 

The index itself is unchanged; pages keep every word. Only the query is focused, which keeps candidate lists centered on content words without affecting what can be found.

3. One row per thing you see

SpotRecall already collapses duplicate content. URLs are normalized (tracking and session parameters removed), and pages with an identical content hash are shown once, keeping the best-ranked copy.

Some pages are near-identical rather than identical. Two sessions of the same chat app, for example, can differ by a few characters in the captured body, which gives them different hashes. On screen they look the same: same title, same preview text.

1.3.1 adds a second, display-level pass. Rows that share both their title and their visible snippet are shown once:

display key = title + "\n" + first 300 characters of the description or text 

This follows how users judge duplicates — by what they can see. As with the content hash, an empty key never merges anything: a record without a title always keeps its own row.

The pairing table

Every recall feature in SpotRecall now has a precision counterpart:

  • Match any word (forgiving for CJK and long phrases) — paired with query-side stopwords.
  • Prefix matching (gar → gardening) — paired with a minimum term length.
  • Bookmark and tab backfill (useful from day one) — paired with the title-coverage prior.
  • URL normalization and content-hash dedupe — paired with display-level dedupe.

Looking at the list this way also gives a design rule for the future. Whenever a change widens what search can find, ask what tells the ranking which of the new candidates are worth showing first.

Scope and limits

  • The stopword list is intentionally tiny and applied to queries only. It is not stemming or a linguistic model.
  • The title-coverage check uses the same tokenizer and prefix rules as search, so it agrees with how matching already works.
  • Display-level dedupe merges rows that look identical. If two genuinely different pages share a title and their opening text, a more specific query still reaches either one.

The takeaway

Search quality is often won in small, explainable signals: which words carry meaning, how much a title-only match really proves, and what a user will perceive as a duplicate. Each rule here is a few lines of code, a test, and a measurement — and together they made the first screen of results noticeably more focused.

SpotRecall 1.3.1 is rolling out on the Chrome Web Store and Microsoft Edge Add-ons, with release notes in 21 languages.

SpotRecall is open source under AGPL-3.0. spotrecall.com · source on GitHub · install from the Chrome Web Store.

This is Post 13 in Building SpotRecall, and the third post in Season 3: sharper, measured search.