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

[nova28]

Adds EmbeddingGemma 2 (Apache 2.0, 768-dim, 8192-token context, multilingual) as two text models backed by onnx-community/embeddinggemma-2-ONNX:

model	file	size
google/embeddinggemma-2	onnx/model.onnx (+ .onnx_data)	1.08 GB
google/embeddinggemma-2-Q	onnx/model_quantized.onnx (+ .onnx_data), int8	0.31 GB

Both are BuiltinSentenceEmbedding models, like google/embeddinggemma-300m: the graph's second output is the pooled, projected and normalized sentence_embedding. Only the text graph is downloaded; the vision and audio encoders are separate files and are not fetched.

Two changes beyond the model entries

1. Empty modality inputs. The text graph also declares image_features, video_features and audio_features inputs ([num_*_tokens, 512]). BuiltinSentenceEmbedding._preprocess_onnx_input now feeds any missing *_features input as a zero-row tensor. It's a no-op for graphs without such inputs, so the existing models are unaffected.
2. Max context length. The export's tokenizer_config.json has transformers' model_max_length sentinel (1e30), which load_tokenizer rightly rejects. I added an optional default_max_length to load_tokenizer / _resolve_max_context. A model class can declare it for exports like this one, and a usable model_max_length / max_length in the config still wins. BuiltinSentenceEmbedding passes 8192 for the two new models only. If you'd rather not grow the load_tokenizer signature, the alternative is fixing model_max_length in the HF repo (or hosting a mirror, as with Qdrant/Qwen3-Embedding-0.6B-onnx); DEFAULT_MAX_LENGTHS could then go.

Canonical values

Computed with the reference implementation: SentenceTransformer("google/embeddinggemma-2", model_kwargs={"dtype": torch.float32}) on sentence-transformers 6.1.0 and transformers 5.19.0, using the same prefixed inputs as the existing EmbeddingGemma tests. Against those values:

	doc hello world max abs diff (first 5)	query max abs diff (first 5)	cosine to reference
google/embeddinggemma-2	< 1e-6	< 1e-6	1.00000
google/embeddinggemma-2-Q	2.8e-4	3.6e-4	0.99996–0.99997

Both models therefore share the reference values in CANONICAL_VECTOR_VALUES / CANONICAL_QUERY_VECTOR_VALUES (within the tests' atol=1e-3).

Tests run locally

• tests/test_preprocessor_utils.py: 43 passed, including the new test_default_max_length_fills_unusable_config (sentinel → default, config key still wins).
• test_embedding / test_query_embedding assertions for google/embeddinggemma-2-Q via a fresh Hugging Face download: pass, truncation.max_length == 8192.
• google/embeddinggemma-2 (fp32), loaded from the same repo files with the unmodified tokenizer_config.json: matches the reference as above.
• ruff check / ruff format --check (v0.3.4, per pre-commit) and mypy fastembed --disallow-incomplete-defs --disallow-untyped-defs --disable-error-code=import-untyped: clean.

The fp32 model is 1.08 GB, so it's skipped by the local size_in_GB > 1 rule and covered by the manual/weekly CI run.

🤖 Generated with Claude Code

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

• Configuration used: Organization UI
• Review profile: CHILL
• Plan: Advanced
• Run ID: 91288028-4729-4a7a-bda8-42f095b6e9a9

📥 Commits

Reviewing files that changed from the base of the PR and between 539499b and d47ea52.

📒 Files selected for processing (4)

• fastembed/common/preprocessor_utils.py
• fastembed/text/builtin_sentence_embedding.py
• tests/test_preprocessor_utils.py
• tests/test_text_onnx_embeddings.py

Included review availability: This review used your included allowance. Your plan provides up to 8 included reviews per hour; 7 remain after this review.

---

📝 Walkthrough

Walkthrough

The tokenizer utilities now accept a default maximum length when tokenizer configuration has no usable limit. The supported model list adds full-precision and quantized google/embeddinggemma-2 entries. Tokenizer loading applies an 8192-token fallback for these models, and ONNX preprocessing supplies zero-row tensors for missing _features inputs. Tests cover tokenizer limit selection and document and query embeddings for both model variants.

Priority: ⬇️ Low

Estimated code review effort: 3 (Moderate) | ~20 minutes

Change: Feature

Merge Risk: ⚪ Minimal · up to d47ea

No actionable issue remains; the change is mergeable after normal checks.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 40.00% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 10 functions across 4 files.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly identifies the main change: adding google/embeddinggemma-2 and its int8 variant.
Description check	✅ Passed	The description directly explains the two new models, their implementation, supporting changes, and validation results.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

• Fix all pre-merge checks with AI

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create a new PR

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}