Google has announced EmbeddingGemma 2, an open embedding model with 740 million parameters. Embeddings are numerical representations of content: a model translates text, images or code into long series of numbers, known as vectors, whose distance from one another reflects semantic similarity. This is precisely what semantic search and recommendation functions are built on – that is, search that works by meaning rather than by exact character string.

What is remarkable about the announcement is less the scope of features than the resource requirements. According to Google, the model requires only around 191 MB of memory and runs directly on the end device. It is also said to outperform models twice its size in some comparative benchmarks. These figures come from the manufacturer and should accordingly be read as its own claims.

Multiple data types in a single vector space

EmbeddingGemma 2 processes not only text but also images, video, audio and code. For projects this means that different content types can be made searchable via the same mechanism, without having to build a separate pipeline for each data type.

Typical fields of application in web and CMS environments are:

  • semantic search across editorial content that also finds synonyms and paraphrases
  • similarity suggestions for related articles, products or documents
  • finding images and media files based on descriptive input
  • matching incoming enquiries to existing knowledge bases

Offline RAG as the real lever

In combination with Gemma 4, Google cites privacy-focused offline RAG applications as a scenario. RAG stands for Retrieval Augmented Generation: a language model does not answer a question from memory but is first supplied with relevant document sections from its own knowledge base. The embedding model handles the search part, while the language model formulates the answer.

Until now, such setups regularly send content to external services – either when generating the vectors or when generating the answer. If both steps can run locally, this outflow is eliminated. For organisations with internal manuals, contracts, HR data or customer documents, this noticeably changes the starting position when assessing matters under data protection law, because no data processing agreement with a cloud provider is needed for these processing steps.

Cost side and operation

In cloud scenarios, embedding calls are an ongoing expense. They are incurred not only once during indexing but with every change to content and with every search query. A model that can be run on ordinary hardware shifts these variable costs into plannable infrastructure.

At the same time, operational effort arises elsewhere. The model must be deployed, versioned and monitored. Changing the model version generally entails recalculating all stored vectors, because vectors from different models are not comparable with one another. Anyone introducing a vector database should plan for this reindexing scenario from the outset.

Assessment for TYPO3 and web projects

For CMS projects the architecture is manageable: when content is saved, text sections are passed to the model and the resulting vectors land in an index. A search query is translated into a vector in the same way and matched against the index. The list of hits can be output directly or passed on as context to a language model.

A step-by-step approach makes practical sense:

  1. Choose a clearly defined use case, such as internal document search.
  2. Clean up the data base – unstructured or outdated content yields weak results even with a good model.
  3. Build a test set of real search queries with expected hits in order to make quality measurable.
  4. Only then consider a generative answer layer.

Managing expectations remains important. A compact model does not replace editorial structuring, and benchmark values say little about hit quality in a specific specialist domain. Whether the results hold up only becomes apparent with your own content and your own vocabulary.

What follows from this

The trend towards small, locally runnable models significantly lowers the barrier to entry for AI features in web and software projects: data protection concerns and ongoing API costs have so far been the two most common reasons for postponing such features. For projects with sensitive content, it is therefore worth planning search and knowledge access now as a separate component instead of tying them firmly to a cloud service. What remains decisive is the effort for operation, index maintenance and quality assurance – this does not disappear but shifts in-house.

Sources