Lovable: Gemini Embedding 2 Becomes Default for Search and Recommendations
LovableLovable now defaults to Gemini Embedding 2 (google/gemini-embedding-2) whenever it builds semantic search, recommendations, or retrieval-augmented generation into a project. Unlike the previous default, Gemini Embedding 001, the new model can embed images, audio, video, and PDFs in addition to text, letting Lovable-built apps search across richer content types. Existing tables that already store Gemini Embedding 001 vectors keep using that model, so nothing breaks for apps already in production.
Key Takeaways
- Gemini Embedding 2 replaces Gemini Embedding 001 as Lovable's default model whenever it builds semantic search, recommendations, or RAG into a project.
- The new model is natively multimodal, embedding images, audio, video, and PDFs in addition to text — the previous default handled text only.
- Lovable preserves backward compatibility by keeping existing tables on Gemini Embedding 001 so already-stored vectors and similarity results aren't disrupted.
- The change unlocks new use cases for Lovable-built apps, such as searching media libraries by visual or audio similarity rather than just text matching.
- The update continues a pattern of Lovable keeping pace with Google's model releases, following earlier default-model updates like Gemini 3.6 Flash and Gemini 3.7 Flash.
- Developers don't need to configure anything — the switch is automatic for any new semantic search, recommendation, or RAG feature Lovable generates going forward.
A New Default Embedding Model
Lovable has switched the default embedding model it reaches for whenever its AI assembles semantic search, recommendation systems, or retrieval-augmented generation (RAG) features into a project. From now on, any new feature of this kind uses Gemini Embedding 2 (google/gemini-embedding-2) instead of the prior default, Gemini Embedding 001.
Why It Matters: Multimodal Embeddings
The headline change is multimodal support. Gemini Embedding 001 could only embed text, which limited semantic search and RAG features built by Lovable to text-based content. Gemini Embedding 2 can embed images, audio, video, and PDFs alongside text, so apps built on Lovable can now support search and recommendation experiences across a much wider range of content — for example, searching a media library by visual similarity, or building RAG over PDF documents instead of only plain text.
Backward Compatibility for Existing Apps
Lovable designed the rollout so existing projects are not disrupted. Tables that already store vectors generated by Gemini Embedding 001 continue to use that model, rather than being silently migrated to a new embedding space, which would break existing similarity search results. Only new tables and new semantic search or RAG features built going forward default to Gemini Embedding 2.
No Extra Setup Required
The switch is automatic. Developers building semantic search, recommendation, or RAG functionality into a Lovable project do not need to configure anything to take advantage of the new model — Lovable selects it by default for any newly generated feature of this kind.