Index Into More Than One Engine

Pass engine when you create each index. The uid is yours to choose; a suffix per engine keeps them apart.

for (const engine of ["meilisearch", "typesense", "elasticsearch", "visor"]) {
  await hb("POST", "/v1/indexes", ADMIN, { uid: `products_${engine}`, primaryKey: "id", engine });
}

Each index goes on your project's engine of that type. If the project has none yet, Headband creates one on first use, within your plan's engines per type.

Send the same array to each index with POST /v1/documents?index=<uid> (or /v1/bulk). Give every document a string id: it works the same on all four, and Typesense hands ids back as strings anyway.

Each engine applies a different part of the settings vocabulary. To make four indexes behave alike, send each the settings it applies, and leave out the ones it refuses.

"Refused" means the settings call answers 400 (unsupported_setting, or feature_not_supported on Visor) and changes nothing. Sending back the settings you read from an index never fails. Two more rules:

• On Typesense, add the documents first. A field can be made sortable only once a document has given it a type; before that the call answers 409 settings_need_documents.
• On Elasticsearch, send none of the attribute lists or synonyms. It searches, filters and sorts every field without them.

const shared = {
  searchableAttributes: ["title", "brand", "category"],
  filterableAttributes: ["brand", "category", "price"],
  sortableAttributes: ["price"],
  synonyms: { coat: ["jacket"] },
};
const SETTINGS = { meilisearch: shared, visor: shared, typesense: shared, elasticsearch: null };

for (const [engine, settings] of Object.entries(SETTINGS)) {
  if (!settings) continue; // Elasticsearch: nothing to send
  await hb("PATCH", `/v1/indexes/products_${engine}/settings`, ADMIN, settings);
}

A search request is the same on every engine, with one exception: on Elasticsearch an exact filter, a facet or a sort on a text field names <field>.keyword. Elasticsearch stores a text field as its analysed words, so brand = "Contoso" finds nothing there (while brand = "contoso" happens to match a lower-cased word). Numbers, booleans and dates need no suffix.

const field = (engine, name) => (engine === "elasticsearch" ? `${name}.keyword` : name);

await hb("POST", "/v1/search", SEARCH, {
  index: `products_${engine}`,
  q: "",
  filter: `${field(engine, "brand")} = "Contoso"`,
  facets: [field(engine, "brand")],
});

The example's output on Meilisearch, Typesense and Visor:

q "running", sort price:asc
  meilisearch   2 hit(s) [1,2]
  typesense     2 hit(s) [1,2]
  visor         2 hit(s) [1,2]

brand = "Contoso", facets brand
  meilisearch   4 hit(s) [1,3,5,8] facets={"brand":{"Contoso":4}}
  typesense     4 hit(s) [8,5,3,1] facets={"brand":{"Contoso":4}}
  visor         4 hit(s) [1,3,5,8] facets={"brand":{"Contoso":4}}

price < 100
  meilisearch   3 hit(s) [6,7,8]
  typesense     3 hit(s) [8,7,6]
  visor         3 hit(s) [6,7,8]

q "coat" (synonym of jacket)
  meilisearch   3 hit(s) [6,4,5]
  typesense     6 hit(s) [6,5,4,8,3,1]
  visor         6 hit(s) [6,4,5,1,3,8]

Filters, facets and a sorted search agree. Without a sort, the order differs. coat finds the three jackets through the synonym on every engine, but Typesense and Visor also match Contoso for it: their typo and prefix matching is looser on short words than Meilisearch's, which allows no typo in a word under 5 letters. On Elasticsearch the facet comes back under brand.keyword.

• Relevance. Typo rules, prefix matching and ranking differ, so the same query can return different hits in a different order. Sort explicitly when the order matters.
• Tasks. Only Meilisearch queues writes. On the others GET /v1/tasks has nothing to wait for.
• Scores. relevanceScore in AI search is set on Meilisearch and Visor only, and hybrid search's _scoring means something different on each engine.
• Search parameters. attributesToCrop, cropLength, matchingStrategy and showMatchesPosition are ignored on Typesense and Elasticsearch, and Visor accepts only its own values. mode is Visor's alone.
• Totals and paging. Visor counts a search's hits exactly only within its first 1000 (past that, totalHitsExact is false). Listing documents by offset stops at 10,000 on Elasticsearch and 1000 on Visor; walk by cursor to read everything.
• Vectors. Hybrid search needs a vector field set up differently on each engine, and Visor has none. See Hybrid Search.

To move one index to another engine rather than keep copies, use an engine switch (POST /v1/indexes/:uid/engine, or headband indexes switch-engine). It keeps the uid and keys, and carries the settings the new engine applies. It lists the rest as drops and refuses until you accept them.