HackerSearch

​

Key features

• The entire corpus: Stories, Ask HN, Show HN, Launch HN, jobs, polls, and comments, all in one index. Two tabs, News and Comments, split the corpus and the facet rail adapts to each.
• Inline autocompletion: As you type, the search box suggests the rest of the current word in grey, taken from the most popular matching story title. Press Tab or → to accept it.
• Disjunctive facet counts: Each keystroke sends a single multi-search request. For every facet with an active selection, it adds a query that excludes that facet’s own filter, so counts stay accurate while you combine filters (disjunctive facets).
• Filters on every dimension: Narrow results by type, time range, minimum points, domain, and author. Domain and author use facet search to find values among thousands of options.
• Relevance, newest, or points: Switch between relevancy ranking and explicit sorts. The sort ranking rule sits before attribute so the Newest and Points modes take priority, and points:desc acts as a final tiebreaker for relevancy.
• Live sync: A Rust indexer reads the Hacker News Firebase API, backfills the full history, then polls for new and updated items every 30 seconds.
• Threaded comments: Open any story or comment to read its full discussion tree, fetched directly from the index.
• Shareable searches: The query, filters, and sort live in the URL, so every search can be shared as a link.

​

How it’s built

• Searchable attributes: title and text only. URLs, domains, and author names are excluded from full-text search to avoid noisy matches (for example, “dan” matching every post by the user “dang”) and remain reachable through filters.
• Granular filterable attributes: Each attribute only enables the features the UI needs. domain and author support facet search, points and created_at support comparison operators, and the rest only use equality.
• Proximity by attribute: proximityPrecision is set to byAttribute, which lowers indexing cost since titles and comment text are independent fields.

​

How faceting stays fast

• High-cardinality facets load on demand: The Domain and Author sections start collapsed. Their names are only added to the facets parameter once you expand them, and a section stays open automatically while it has a selected value. The low-cardinality tags facet (story, Ask HN, Show HN, and so on) is always computed on the News tab and skipped entirely on the Comments tab.
• Exclusion queries only for active selections: Disjunctive faceting needs a separate query per facet, with that facet’s own filter removed. HackerSearch only sends that query for facets where you have selected a value. For the others, the counts from the main query are already correct.
• Facet search instead of long lists: Each expanded section shows its top values. Typing in its filter box calls the facet search endpoint to search across all values, so the page never has to load full distributions. This is why only domain and author enable facetSearch in the index settings.
• Cancel superseded requests: When you keep typing, the in-flight request for the previous query is aborted, so slow responses never pile up.

​

How autocompletion works

{
  "indexUid": "hn",
  "q": "meil",
  "attributesToSearchOn": ["title"],
  "filter": "type != \"comment\"",
  "sort": ["points:desc"],
  "limit": 1,
  "attributesToRetrieve": ["title"]
}

• attributesToSearchOn limits matching to story titles, so suggestions read like real headlines rather than fragments of comments. See configure searchable attributes.
• filter excludes comments, which have no title.
• sort: ["points:desc"] returns the most upvoted matching story. Meilisearch treats the last query word as a prefix, so meil already matches “Meilisearch”.
• limit: 1 and attributesToRetrieve keep the response tiny, since only one title is needed.

​

Links