In brief
Cloudflare’s 12 November update brings EmbeddingGemma 2 to Workers AI. Start with an approved product collection, compare multilingual search results with known answers, and test access controls before expanding.
What Cloudflare announced
On 12 November, Cloudflare announced that Google DeepMind’s EmbeddingGemma 2 is available on Workers AI. The model turns text, images, video and audio into numerical representations called embeddings, designed to make related content searchable together.
Cloudflare says the model supports more than 100 languages and places these inputs in a shared 768-dimensional space. These are vendor-described capabilities, not results from a MIRA CODE client deployment or an independent benchmark.
A useful starting point: one product collection
For an SME, a practical experiment could pair a small product catalogue with approved product photos and extracted text from manuals. Staff could try everyday searches in English, Malay and Chinese, including descriptions that do not match the exact product name.
For example, a lighting retailer might test whether “small bedside lamp” retrieves a suitable product photo and the relevant specification sheet. This is a proposed use case, not a promise that every query will work. Document text still needs to be prepared, and the application needs an index, a search interface and links back to the original records.
Our recommendation: measure wrong answers too
Prepare a reviewed set of queries with expected results, including similar-looking products, local terminology and questions with no valid match. Compare the first few results with the existing search rather than judging a handful of impressive examples.
Keep exact product codes and specification filters available. A visually similar item can have a different size, voltage or compatibility. Show the source document and product identifier so staff can verify the match. Record false matches, missed results, response time and the total cost of indexing and serving searches.
Keep permissions attached to the content
Before indexing, confirm which materials the business is permitted to send to the provider. Separate public catalogue content from customer records and restricted internal documents.
Enforce access rules before returning results, including titles, snippets and thumbnails. Test with accounts that have different permissions, and check that removed or outdated documents stop appearing. An embedding model does not supply these application-level controls.
A small, reviewed pilot can show whether this approach improves a real workflow. Explore MIRA CODE’s AI automation and web development services below to discuss the surrounding application and search experience.
