Skip to main content

​
Generating a user profile

Search personalization requires a profile of the user performing the search. Meilisearch does not yet provide automated generation of user profiles. You’ll need to dynamically generate a user profile for each search request. This should summarize relevant traits, such as:
  • Category preferences, like brand or size
  • Price sensitivity, like budget-conscious
  • Possible use cases, such as fitness and sport
  • Other assorted information, such as general interests or location
The re-ranking model only processes positive signals. It cannot interpret negative statements like “dislikes blue” or “is not interested in luxury brands”. Always use affirmatively stated preferences instead: “likes the color red”, “prefers cheaper brands”. Once search personalization is active and you have a pipeline in place to generate user profiles, you are ready to perform personalized searches. Submit a search query and include the personalize search parameter. personalize must be an object with a single field, userContext. Use the profile you generated in the previous step as the value for userContext: 
curl \
  -X POST 'MEILISEARCH_URL/indexes/INDEX_NAME/search' \
  -H 'Content-Type: application/json' \
  --data-binary '{
    "q": "wireless keyboard",
    "personalize": {
      "userContext": "The user prefers compact mechanical keyboards from Keychron or Logitech, with a mid-range budget and quiet keys for remote work."
    }
  }'

​
Reranking scope and the limit parameter

The reranker only sees the documents that Meilisearch returns for a given query. This means limit directly controls how many documents the reranker can choose from. If you set limit to 20, the reranker picks the best order from those 20 documents. It cannot promote a document that Meilisearch did not return. Increasing limit gives the reranker a wider pool to work with, but also increases latency and the number of tokens sent to the reranking model. A limit between 20 and 100 is a reasonable starting point for most personalized search use cases. Go higher if ranking quality matters more than latency for your application.

​
Limits

Personalized search is subject to the following Cohere reranking limits:
  • userContext + q combined must stay under 2048 tokens
  • Each document sent to the reranker can be at most 32,664 tokens
  • The reranker accepts at most 10,000 documents per request, so limit cannot usefully exceed 10,000
As a rough guide, one word is about 2 to 3 tokens and a paragraph is about 128 tokens. When q is empty, Meilisearch returns a deterministic set of documents based on ranking rules and filters. The userContext can only reorder that fixed set. It cannot change which documents are retrieved. This has two consequences worth understanding before building a personalized feed:
  • Two very different userContext values will return the same documents, just in a different order, because the underlying candidate pool is the same for both calls
  • Two slightly different phrasings of the same userContext can produce different result orders, because the reranker is sensitive to wording
If you need a genuinely personalized candidate set for a homepage feed or infinite scroll, consider these alternatives:
  • Use userContext as q with full semantic search. Set semanticRatio: 1 and pass the userContext string directly as q. In this case, the personalize parameter is not needed: the semantic retrieval already personalizes which documents are returned. Reserve the personalize parameter for when the user has typed their own query.
  • Run one search per interest area. Generate a short list of distinct interest terms from the user profile and run a multi-search with one semantic query per term. Each query gets its own retrieval pass, so the merged result pool naturally covers more variety. This produces more diverse results but uses more queries.
  • Increase limit. If you keep q empty, a higher limit gives the reranker a wider fixed pool to reorder within. This is the lowest-effort option but does not change which documents are retrieved.

​
Next steps

Generate user context

Build dynamic user profiles for more relevant personalized results.

Personalize e-commerce search

Apply search personalization to an e-commerce product catalog.

Recommendations

Build a recommendation system with the similar documents endpoint.