Releases 13.1 

Key Highlights 

1. Initial Vector Search Introduction:

   • The EXT:solr version 13.1 introduces an initial vector search option as a new search variant.
   • Activating this feature automatically generates vectors during indexing and frontend search.
   • A connected large language model (LLM) is required for operation, though it is not directly related to EXT:solr.

2. Handling and Limitations of Vector Search:

   • Current implementation includes limitations, especially in error handling when required LLMs are not defined, leading to potential impairments in indexing or search.
   • Indexing without vector calculation results in documents not being found despite successful index status.
   • A missing or unavailable LLM during search attempts can lead to a SolrInternalServerErrorException, returning an HTTP status 500.

3. Configuring a Large Language Model:

   • To use vector search, a large language model must be connected to encode text into vectors.
   • Configuration details are available in the Apache Solr Reference Guide 9.9.

   • Models can be uploaded using a JSON file and cURL command.

     Example configuration for the JSON file

     {
       "class": "dev.langchain4j.model.openai.OpenAiEmbeddingModel",
       "name": "llm",
       "params": {
         "baseUrl": "https://api.openai.com/v1",
         "apiKey": "apiKey-openAI",
         "modelName": "text-embedding-3-small",
         "timeout": 5,
         "logRequests": true,
         "logResponses": true,
         "maxRetries": 2
       }
     }

     Copied!
   • The number of dimensions for vectors defaults to 768 but can be adjusted via the SOLR_VECTOR_DIMENSION environment variable.

Future Developments 

• Improved Error Handling: Future versions plan to enhance error handling during vector indexing and search to increase robustness and reliability.
• Additional Query Types: New query types such as vector sorting and vector re-ranking are planned, allowing for more advanced search result manipulation.
• Backend Module for LLM Management: A backend module for managing large language models is anticipated, simplifying maintenance and configuration for developers.

This introduction marks a significant advancement for TYPO3's search capabilities by integrating AI technologies, with ongoing improvements and features planned for future releases.

For more detailed technical implementation and setup instructions, users should refer to the version 13.1 release notes and the associated documentation sections.

Technical insights 

As soon as vector search is enabled, EXT:solr will use the connected LLM to generate vectors during indexing and for each search in the frontend. During indexing vectors are generated based on field vectorContent which is by default filled with the contents of the content field. TypoScript indexing configurations can be used to customize the contents of the vectorContent field, e.g.:

plugin.tx_solr.index.queue.news.fields {
  vectorContent = SOLR_CONTENT
  vectorContent.cObject = COA
  vectorContent.cObject {
    10 = TEXT
    10 {
      field = name
    }

    15 = TEXT
    15 {
      field = bodytext
    }
  }
}

During indexing vectors will be created and stored in field vector.