For the complete documentation index, see llms.txt. This page is also available as Markdown.

Hybrid Vector Search

Merchants with Hybrid Vector Search enabled gain the ability to search products based on conceptual similarity in addition to precise retrieval based on textual relevance.

Depending on merchant preferences, Hybrid Vector Search can take over from keyword search when:

  • The search query has a poor click-through rate.

  • The search query has a poor conversion rate.

  • The search query has no results.

  • Hybrid Vector Search was requested specifically for a particular search query.

In principle, Hybrid Vector Search works out of the box and doesn't require changes to the integration. The suggestions below are optional.

Recognizing the type of search logic being used

To understand results, it's valuable to know which type of logic generated them. The search API exposes this information in the SearchProducts object of the search response.

searchType contains the type of search logic being used:

  • keyword indicates pure keyword search.

  • vector indicates pure vector search.

  • hybrid indicates hybrid vector search (blended keyword + vector search).

This information can optionally be used in the search result page to convey whether these results are precise (keyword search) or conceptually related (Hybrid Vector Search). Hybrid Vector Search results can be more general than keyword search results — communicating this to the user helps with setting expectations.

The response also provides searchTypeReason which indicates why a particular searchType was used. Most notably, this will list the hybrid vector search trigger that was applied e.g., 0 keyword search results, low CTR.

Tracking requirements

To enable Hybrid Vector Search analytics, include searchType and searchTypeReason in impression and click tracking events sent to Nosto.

The searchType and searchTypeReason values queried from the search API can be included in tracking events verbatim within the tracking metadata properties searchType and searchTypeReason, on the same level as query. Here is an example for a well-formed tracking metadata that includes both:

When using NostoJS' automatic tracking (using the track parameter), searchType and searchTypeReason tracking is included automatically, and no adjustments to the integration are required.

API integrations and integrations managing tracking metadata manually need to take care of passing through both searchType and searchTypeReason explicitly.

Limitations

When Hybrid Vector Search engages, features of Nosto search work as normal with one major caveat: Only the 1000 most relevant results are accessible via pagination and covered by facets and sorting. Depending on the strictness of the relevance threshold defined in the Hybrid Vector Search settings, fewer results could be available.

This cutoff is based purely on Hybrid Vector Search relevance and does not take pinning or promote/demote rules into account.

Last updated

Was this helpful?