EmbeddingGemma 2

EmbeddingGemma 2 (google/embeddinggemma-2) embeds text, images, audio, video, and ordered combinations of them into one vector space. Use it when a single index should hold mixed media, or when you want one embedding per document rather than one per part. Antfly runs it on CPU, Metal, and CUDA (Ampere or newer).

Pull the model#

antfly inference pull hf:google/embeddinggemma-2:safetensors@914f7f89142e33e77833254d9c9b90c3cef7303b --tasks embed

The weights keep their upstream license.

Embed#

Flat inputs return one vector per text or media part. A group returns one vector for all of its ordered content. Post this to /ai/v1/embed:

{
  "model": "embeddinggemma2",
  "task_type": "RETRIEVAL_DOCUMENT",
  "dimensions": 256,
  "input": [
    {
      "title": "Account recovery",
      "content": [
        {"type": "text", "text": "How to reset a forgotten password."},
        {"type": "image_url", "image_url": {"url": "https://example.com/reset.png"}}
      ]
    }
  ]
}

• dimensions is 768 (default), 512, 256, or 128.
• task_type is RETRIEVAL_DOCUMENT (default), RETRIEVAL_QUERY, QUESTION_ANSWERING, FACT_VERIFICATION, CODE_RETRIEVAL_QUERY, SEMANTIC_SIMILARITY, CLASSIFICATION, or CLUSTERING. Index with the document role and search with the query role; Antfly does this automatically for index embedders.
• title applies only to RETRIEVAL_DOCUMENT groups that contain text.
• A request holds up to 64 groups of up to 64 parts, and a group must fit 8192 tokens including expanded media. Inputs that overflow fail rather than truncate.
• Video (MP4, MOV, or WebM) goes in a group as a media part and is sampled at 1 FPS, up to 32 frames. CUDA video needs Linux NVDEC and an H.264 MP4 of at most 60 seconds.

From the CLI:

antfly inference embed models/embedders/embeddinggemma2 --group \
  --title 'Account recovery' --dimensions 256 \
  --text 'Reset a forgotten password.' --image reset.png

Pin the model in an index#

Each response includes a model_identity hash for the exact weights, processor, and execution recipe. Set it as the index embedder's model_identity so that a different model generation is rejected instead of silently mixing vectors. Rebuild the index when you change models, dimensions, or task profile.

Similarity decisions#

EmbeddingGemma 2 can answer choice and multi_choice questions on /ai/v1/decisions with decision_method: "embedding_similarity": each choice is a description or a few examples, and the answer reports cosine similarities and a margin. Scores are similarities, not probabilities; set min_similarity, min_margin, or similarity_thresholds from your own evaluation. See Typed Decisions for the request format, batches, and SQL use.

Qualification evidence and benchmarks are in the EmbeddingGemma 2 history.