Skip to main content
NeuralSearch uses a different infrastructure than Algolia’s keyword search, so some features work differently or aren’t supported. Review these limitations before you activate NeuralSearch or apply it to production traffic.

​
Virtual replicas

NeuralSearch doesn’t support virtual replicas, which are used for relevant sorting. To use NeuralSearch with sorted results, use a standard replica instead of a virtual replica. If you want custom ranking criteria to affect vector results, enable usePositionalSemanticRanking. For more information, see Configure NeuralSearch with the semantic settings endpoint.

​
Natural language processing (NLP) features

NLP features apply only to the keyword search part of NeuralSearch. Semantic (vector) retrieval doesn’t use these features:
  • Synonyms. Synonyms configured in your index expand keyword queries but don’t affect the vector representation used for semantic retrieval.
  • Decompounding (used for German and other compound-word languages) applies only to keyword matching.
  • Pluralization and word inflections. Stemming and plural normalization apply to keyword matching. The semantic model handles linguistic variation through embeddings but explicit rules configured in Algolia only affect keyword retrieval.
If you need precise NLP behavior, keyword results in NeuralSearch work the same way as keyword-only search. Semantic results supplement keyword results rather than replacing them.

​
Case sensitivity and query normalization

Semantic (vector) retrieval doesn’t normalize queries. Differences in capitalization, spelling, and special characters may produce different semantic results or rankings. For example, table, Table, and TABLE might produce different results. The same applies to canapé and canape. Because NeuralSearch merges keyword and semantic results, these differences can change the final ranking. The keyword part of the query still applies your index’s normalization settings, so this behavior only affects the semantic contribution to the results.

​
Sorting by filter or geographical distance

NeuralSearch supports sort-by replicas for sorting results by a fixed ranking order. Standard sort-by settings don’t apply filter-based ranking or geographic distance to vector-retrieved results by default. To apply geographic distance and filter-based ranking to vector results, enable usePositionalSemanticRanking. For more information, see Configure NeuralSearch with the semantic settings endpoint.

​
Index size limits

NeuralSearch indexes follow the same size limits as other Algolia indexes:
  • Maximum index size: 100 GB
  • Maximum record size: 100 KB
Indexes above these limits aren’t supported, whether they use keyword search or NeuralSearch.

​
Other considerations

  • Custom dictionaries. NeuralSearch doesn’t support custom dictionaries for stop words, plurals, or compounds.
  • Log IP forwarding. NeuralSearch indices don’t show user IP addresses from X-Forwarded-For in search logs.
Last modified on July 24, 2026