new: add google/embeddinggemma-2 and its int8 variant

fastembed/common/preprocessor_utils.py

@@ -47,12 +47,15 @@ def _valid_context(value: Any) -> int | None:
     return value
 
 
-def _resolve_max_context(tokenizer_config: dict[str, Any], model_dir: Path) -> int:
+def _resolve_max_context(
+    tokenizer_config: dict[str, Any], model_dir: Path, default_max_length: int | None = None
+) -> int:
     """Pick the truncation limit, preferring the stricter of the two tokenizer config keys.
 
     `config.json:max_position_embeddings` deliberately is not used as a fallback: it is the size
     of the position table, not the usable context, and the two differ per architecture, e.g.
-    roberta reports 514 for a usable 512.
+    roberta reports 514 for a usable 512. `default_max_length` is a limit the model class
+    declares for an export whose tokenizer config carries none; the config keys still win.
     """
     candidates = [
         context
@@ -62,6 +65,8 @@ def _resolve_max_context(tokenizer_config: dict[str, Any], model_dir: Path) -> i
         )
         if context is not None
     ]
+    if not candidates and default_max_length is not None:
+        candidates = [default_max_length]
     if not candidates:
         raise ValueError(
             f"Could not determine the maximum context length for {model_dir}. Set a positive "
@@ -71,7 +76,9 @@ def _resolve_max_context(tokenizer_config: dict[str, Any], model_dir: Path) -> i
     return min(candidates)
 
 
-def load_tokenizer(model_dir: Path) -> tuple[Tokenizer, dict[str, int]]:
+def load_tokenizer(
+    model_dir: Path, default_max_length: int | None = None
+) -> tuple[Tokenizer, dict[str, int]]:
     tokenizer_path = model_dir / "tokenizer.json"
     if not tokenizer_path.exists():
         raise ValueError(f"Could not find tokenizer.json in {model_dir}")
@@ -90,7 +97,7 @@ def load_tokenizer(model_dir: Path) -> tuple[Tokenizer, dict[str, int]]:
     with open(str(tokenizer_config_path)) as tokenizer_config_file:
         tokenizer_config = json.load(tokenizer_config_file)
 
-    max_context = _resolve_max_context(tokenizer_config, model_dir)
+    max_context = _resolve_max_context(tokenizer_config, model_dir, default_max_length)
 
     tokens_map = load_special_tokens(model_dir)
 

fastembed/text/builtin_sentence_embedding.py

@@ -1,11 +1,21 @@
+from pathlib import Path
 from typing import Any, Iterable, Type
 
+import numpy as np
 
 from fastembed.common.types import NumpyArray
 from fastembed.common.onnx_model import OnnxOutputContext
+from fastembed.common.preprocessor_utils import load_tokenizer
 from fastembed.text.onnx_embedding import OnnxTextEmbedding, OnnxTextEmbeddingWorker
 from fastembed.common.model_description import DenseModelDescription, ModelSource
 
+# Context length for exports whose tokenizer_config.json carries transformers' "unknown"
+# sentinel instead of a usable `model_max_length`.
+DEFAULT_MAX_LENGTHS: dict[str, int] = {
+    "google/embeddinggemma-2": 8192,
+    "google/embeddinggemma-2-q": 8192,
+}
+
 
 supported_builtin_sentence_embedding_models: list[DenseModelDescription] = [
     DenseModelDescription(
@@ -24,6 +34,39 @@
         model_file="onnx/model.onnx",
         additional_files=["onnx/model.onnx_data"],
     ),
+    DenseModelDescription(
+        model="google/embeddinggemma-2",
+        dim=768,
+        description=(
+            "Text embeddings, Multimodal model used text-only, multilingual, 8192 input tokens "
+            "truncation, Prefixes for queries/documents: `task: search result | query: {content}` "
+            "for query, `title: {title | 'none'} | text: {content}` for documents, 2026 year."
+        ),
+        license="apache-2.0",
+        size_in_GB=1.08,
+        sources=ModelSource(
+            hf="onnx-community/embeddinggemma-2-ONNX",
+        ),
+        model_file="onnx/model.onnx",
+        additional_files=["onnx/model.onnx_data"],
+    ),
+    DenseModelDescription(
+        model="google/embeddinggemma-2-Q",
+        dim=768,
+        description=(
+            "Text embeddings, Multimodal model used text-only, multilingual, 8192 input tokens "
+            "truncation, Prefixes for queries/documents: `task: search result | query: {content}` "
+            "for query, `title: {title | 'none'} | text: {content}` for documents, int8 weights, "
+            "2026 year."
+        ),
+        license="apache-2.0",
+        size_in_GB=0.31,
+        sources=ModelSource(
+            hf="onnx-community/embeddinggemma-2-ONNX",
+        ),
+        model_file="onnx/model_quantized.onnx",
+        additional_files=["onnx/model_quantized.onnx_data"],
+    ),
     DenseModelDescription(
         model="ibm-granite/granite-embedding-small-english-r2",
         dim=384,
@@ -58,6 +101,27 @@ def _list_supported_models(cls) -> list[DenseModelDescription]:
         """
         return supported_builtin_sentence_embedding_models
 
+    def _load_tokenizer(self, model_dir: Path) -> None:
+        self.tokenizer, self.special_token_to_id = load_tokenizer(
+            model_dir=model_dir,
+            default_max_length=DEFAULT_MAX_LENGTHS.get(self.model_name.lower()),
+        )
+
+    def _preprocess_onnx_input(
+        self, onnx_input: dict[str, NumpyArray], **kwargs: Any
+    ) -> dict[str, NumpyArray]:
+        """Feed empty modality inputs to multimodal graphs used for text only.
+
+        The embeddinggemma-2 export is a single graph that also declares `image_features`,
+        `video_features` and `audio_features` inputs. For text, each is a zero-row tensor.
+        """
+        for node in self.model.get_inputs():  # type: ignore[union-attr]
+            if node.name in onnx_input or not node.name.endswith("_features"):
+                continue
+            width = node.shape[-1] if isinstance(node.shape[-1], int) else 0
+            onnx_input[node.name] = np.zeros((0, width), dtype=np.float32)
+        return onnx_input
+
     def _post_process_onnx_output(
         self, output: OnnxOutputContext, **kwargs: Any
     ) -> Iterable[NumpyArray]:

tests/test_preprocessor_utils.py

@@ -279,6 +279,27 @@ def test_unusable_max_context_raises(make_model_dir, model_max_length, max_lengt
         load_tokenizer(model_dir)
 
 
+@pytest.mark.parametrize(
+    "model_max_length,max_length,expected",
+    [
+        (HF_SENTINEL, None, 8192),  # onnx-community/embeddinggemma-2-ONNX
+        (None, None, 8192),
+        (512, None, 512),  # a usable config key still wins over the declared default
+        (HF_SENTINEL, 256, 256),
+    ],
+)
+def test_default_max_length_fills_unusable_config(
+    make_model_dir, model_max_length, max_length, expected
+) -> None:
+    model_dir = make_model_dir(
+        tokenizer_config={"model_max_length": model_max_length, "max_length": max_length},
+    )
+
+    tokenizer, _ = load_tokenizer(model_dir, default_max_length=8192)
+
+    assert tokenizer.truncation["max_length"] == expected
+
+
 def test_absent_max_context_keys_raise(make_model_dir) -> None:
     model_dir = make_model_dir(
         drop_from_tokenizer_config=("model_max_length", "max_length"),

tests/test_text_onnx_embeddings.py

@@ -78,6 +78,12 @@
     "google/embeddinggemma-300m": np.array(
         [-0.08181356, 0.0214127, 0.05120273, -0.03690156, -0.0254504]
     ),
+    "google/embeddinggemma-2": np.array(
+        [-0.03013361, 0.01577923, 0.05849417, -0.00649917, -0.03439925]
+    ),
+    "google/embeddinggemma-2-Q": np.array(
+        [-0.03013361, 0.01577923, 0.05849417, -0.00649917, -0.03439925]
+    ),
     "Qwen/Qwen3-Embedding-0.6B": np.array(
         [-0.01476084, 0.01723184, -0.01195498, -0.07275258, 0.00281229]
     ),
@@ -113,16 +119,26 @@
 
 DOC_PREFIXES = {
     "google/embeddinggemma-300m": "title: none | text: ",
+    "google/embeddinggemma-2": "title: none | text: ",
+    "google/embeddinggemma-2-Q": "title: none | text: ",
 }
 QUERY_PREFIXES = {
     "google/embeddinggemma-300m": "task: search result | query: ",
+    "google/embeddinggemma-2": "task: search result | query: ",
+    "google/embeddinggemma-2-Q": "task: search result | query: ",
     "Qwen/Qwen3-Embedding-0.6B": QWEN3_INSTRUCT_PREFIX,
     "Qwen/Qwen3-Embedding-0.6B-Q": QWEN3_INSTRUCT_PREFIX,
 }
 CANONICAL_QUERY_VECTOR_VALUES = {
     "google/embeddinggemma-300m": np.array(
         [-0.22990295, 0.03311195, 0.04290345, -0.03558498, -0.01399477]
     ),
+    "google/embeddinggemma-2": np.array(
+        [-0.02504013, 0.05100445, 0.05460444, -0.02423813, -0.04161748]
+    ),
+    "google/embeddinggemma-2-Q": np.array(
+        [-0.02504013, 0.05100445, 0.05460444, -0.02423813, -0.04161748]
+    ),
     "Qwen/Qwen3-Embedding-0.6B": np.array(
         [-0.01908712, 0.01635596, -0.00356586, -0.03947155, -0.01387356]
     ),