Gemini Provider 1.1.0 Adds Image Embeddings for Drupal AI Search and RAG

Extending semantic retrieval beyond text, Gemini Provider 1.1.0 adds image embeddings to Drupal AI, allowing an image or an image paired with text to be converted into a vector for search and retrieval-augmented generation. Lekso Surameli, a Drupal developer at Omedia, authored the release post published on 29 September 2026. Development of the provider is sponsored by Pantheon.

The new capability uses Drupal AI's existing embeddings operation rather than a separate Gemini-specific workflow. Modules built on that operation can therefore request vectors from images or combined image-and-text inputs. The release documentation identifies use cases including semantic media-library search, finding visually similar products or assets, and indexing image-heavy catalogues or galleries.

Gemini Provider previously supported text-only embeddings. Version 1.1.0 allows an image to be embedded on its own or together with text such as a caption, producing one vector for the combined input. The release also corrects the embedding metadata reported to Drupal AI, including model input limits and vector dimensions.

The gemini-embedding-2 model now reports an input limit of 8,192 tokens and a vector size of 3,072. Older text-only embedding models report a 2,048-token input limit. Those values matter because vector databases need the correct dimensions, while Drupal AI Search can use the provider's reported input limit when determining chunk sizes.

That change creates an upgrade consideration for sites already using gemini-embedding-2. When AI Search relies on the provider's default chunk size, content should be reindexed because the reported limit rises from 2,048 to 8,192 tokens. Sites with an explicitly configured chunk size are unaffected.

New Installs Move to Current Gemini Models

Version 1.1.0 also changes the default models used by new installations after Google retired or restricted models configured by the 1.0 series. Existing sites keep their current settings after updating, so administrators should review the models selected under Drupal AI rather than assuming the new defaults have been applied automatically.

The release notes identify gemini-2.0-flash as a previous speech-to-text default that the Gemini API no longer serves. Existing sites may also retain gemini-2.5-flash for chat, which Google no longer makes available to new API users.

Streaming and File Handling Also Change

Gemini Provider now declares chat streaming as a supported capability. CKEditor's AI assistant can therefore display Gemini responses as they are generated instead of waiting for the complete response.

File handling has also been revised across chat, embeddings and speech-to-text operations. WAV and AIFF recordings using Drupal's MIME type labels can now be submitted for transcription, while images labelled image/jpg are accepted. Files the Gemini client library cannot process, including GIF images and M4A audio, now return an AiResponseErrorException that Drupal AI can log and pass to failover subscribers instead of ending with an uncaught error.

Gemini Provider 1.1.0 supports Drupal 10 and 11 and requires Drupal AI 1.2 or later. The release includes contributions from Surameli, Artem Dmitriiev, Juan Correa Solórzano, Marcus Johansson and István Csáki. Full technical changes and upgrade guidance are available in the Gemini Provider 1.1.0 release notes.