Vector search limitations

Below you can find all limitations and unsupported functionalities when vector search is included:

User experience tuning limitations

Unsupported functionalities

  • Admin Relevancy

  • Personalized User Query

  • Graph Boosting

  • Learn To Rank Results Boosting

  • Learn To Rank Results Suggestion

  • NLQ Processing

  • Similar Documents Query Processor

  • Similar Documents Results Processor

Supported with limited functionality

  • Synonym Processor: Synonyms will continue to work on keyword search. The vector search query must be generated for the original query, before synonym expansion.

  • Refinement Mapper: Refiners will be affected by vector search limitations such as inaccurate refiner counts.

  • Range Facets: Refiners will be affected by vector search limitations such as inaccurate refiner counts.

  • Document Expertise Search: affected by vector search limitations such as inaccurate refiner counts.

Feature limitations

Unsupported functionalities

  • Personalization

  • Landing page widgets

  • Refiners with trustable counts or search within

  • NLQ / Questions and Answers

  • Advanced Search

  • Search Inside

  • Complex KQL queries / Boolean queries

  • Proximity Slider

  • Sorting / Order By

  • Similar Documents

  • Expertise Search

  • SpellCheck

  • Analytics - NOT Available by not using our UI

  • Federating Mixed Vectors / Non-Vector Azure Search Indexes / Differently configured Azure Vector Search Indexes (i.e. different LLMs, different Vector Search field sizes, etc.)

Available with limited functionality

  • Federating multiple Azure Indexes with or without vector fields when vector search is enabled works, but with the following limitations:

    • Additional attention will be needed to normalize the scores via scripting or the score weighting mechanism provided by Smart Hub.

    • Federating multiple indexes, all configured with Vector Search, is possible for indexes with the same vector configuration (i.e. same embeddings providers, same vector fields setup etc.).

    • Pagination works up to page 100 with the risk of lack of performance for high page numbers.

  • Keywords Highlighting functionality has the following limitations:

    • Highlight is generated / applied for the keyword query only.

    • Nothing for the vector part matches will be visible as highlights.

  • Smart Previews module has the following limitations:

    • Do NOT generate embeddings for Preview Validation Queries. To achieve this, modify the query scripting stage that generates the embeddings by adding a query validation. If the query has the format below, it means that the query is performed by the SmartPreviews module and you need to skip generating the vector embeddings.

      "Path=\"https://path1" OR escbasecrawlurl:\"https://path1" OR Path=\"https://path2" OR escbasecrawlurl:\"https://path2" OR ..."

    • All documents returned by the vector search that are eligible for preview generation can be previewed.

    • Highlighting in Smart Preview won’t work as expected. There will be scenarios when no keywords will be highlighted.

Important Mentions

The decision on whether to generate or not vector embeddings to run hybrid search based on documentation and vector search limitations is taken in the pro-serve built scripting component that generates the embeddings.

BA Insight recommends the following:

  • detect that the executed query is complex KQL (i.e. Boolean queries, XRANK queries, properties restrictions queries etc.) and not generate vector embeddings for it.

  • detect that the executed query is a Smart Preview validation query and not generate vector embeddings for it.

  • detect that the executed query is an Expertise query and not generate vector embeddings for it.

  • detect that the executed query is a wildcard query and not generate vector embeddings for it.

  • ensure that the vector query is not generated from the query that was expanded with synonyms but from the original input query.