Tuning Algolia Typo Tolerance for Precise Search Results

The problem: noisy results on exact identifiers

Imagine a product catalog where users search by SKU, model number, or year. With Algolia’s default typo tolerance, a query like ABCD1234 also matches ABCD1235 because the engine treats the trailing digit as a one‑character edit. The result set suddenly contains unrelated items, click‑through drops, and the search feels “broken.”

The fix isn’t to turn typo tolerance off globally — that would hurt the normal fuzzy experience for free‑text queries. Instead, you can scope the behavior: keep fuzzy matching for titles and descriptions, but require exact matches on identifier fields and numeric tokens.

How Algolia applies typos and prefixes by default

• Typo tolerance uses a Levenshtein‑distance algorithm. One typo is allowed for words ≥ 4 characters; two typos for words ≥ 8 characters. Thresholds are controlled by minWordSizefor1Typo and minWordSizefor2Typos.
• Prefix matching is always active on the last query token (type‑as‑you‑go). Earlier tokens must match exactly or via typo tolerance.
• Ranking places the typo criterion after words, filters, and before proximity, so exact matches outrank 1‑typo matches, which outrank 2‑typo matches.

Key configurable parameters

Parameter	Scope	Effect
minWordSizefor1Typo	Index setting	Minimum word length before a single typo is considered (default 4).
minWordSizefor2Typos	Index setting	Minimum word length before two typos are allowed (default 8).
disableTypoToleranceOnAttributes	Index setting (array of attribute names)	Turns off typo tolerance for the listed attributes only.
disableTypoToleranceOnWords	Index setting (array of words)	Prevents typo expansion for specific tokens such as brand names.
allowTyposOnNumericTokens	Index setting (boolean)	When false, numeric tokens (SKUs, years, model numbers) are matched exactly.

Worked example: protecting a SKU field

Assume an index products with records containing title, description, and sku. We want fuzzy matching on title and description but exact matching on sku and any numeric token.

curl -X PUT "https://YOUR_APP_ID.algolia.net/1/indexes/products/settings" \
  -H "X-Algolia-API-Key: YOUR_ADMIN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{\n  \"minWordSizefor1Typo\": 4,\n  \"minWordSizefor2Typos\": 8,\n  \"disableTypoToleranceOnAttributes\": [\"sku\"],\n  \"allowTyposOnNumericTokens\": false\n}'

Where to run: any terminal with network access to Algolia’s API. Permissions: an API key with settings ACL (admin key or a dedicated key). Placeholders: replace YOUR_APP_ID, YOUR_ADMIN_API_KEY, and products with your actual values.

Expected check: after the call, query the index with a one‑character SKU typo:

curl -X POST "https://YOUR_APP_ID.algolia.net/1/indexes/products/query" \
  -H "X-Algolia-API-Key: YOUR_SEARCH_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{\"query\": \"ABCD1235\"}'

Inspect the response’s _rankingInfo.typo field for the record that contains sku: \"ABCD1234\". It should be absent (or have typo: 0 only if the match came from another attribute). If the record still appears with typo: 1, the attribute‑level disable did not take effect — some plans require a re‑index for the disableTypoToleranceOnAttributes setting.

Trade‑offs and verification

• Noise vs. recall: Lowering minWordSizefor1Typo to 1 lets the engine correct typos on very short words (e.g., “a” → “i”), but it floods results on short queries. Keep the default or raise it for cleaner results.
• Re‑index requirement: Changing disableTypoToleranceOnAttributes or disableTypoToleranceOnWords may need a full re‑index on certain pricing tiers. Verify in your dashboard’s “Indices → Settings → Advanced” whether the setting shows “live” or “requires reindex”.
• Interaction with synonyms and rules: A synonym expanded from a typo‑corrected term can cascade into unexpected matches. Test combined behavior in a staging index before promoting to production.
• Monitoring: Export query logs via the Analytics API (or the dashboard) and look at the typo count per query. High typo counts paired with low click‑through signal that the current thresholds are too permissive for that query pattern.

Actionable next steps

1. Identify every attribute that must match exactly (SKU, model, year, ISBN, etc.).
2. Add those attribute names to disableTypoToleranceOnAttributes and set allowTyposOnNumericTokens to false.
3. If you see many one‑character typos on short words, raise minWordSizefor1Typo to 5 or 6.
4. Run the verification queries shown above; confirm the typo field behaves as expected.
5. Schedule a weekly review of the Analytics typo‑rate report to catch regressions early.

By scoping typo tolerance to the fields that truly benefit from fuzziness, you keep the fast, forgiving experience for natural‑language searches while guaranteeing precision on the identifiers that drive purchases and support lookups.