Jarvis Embedding — EmbeddingGemma 300M BF16 GGUF
This repository is the disaster-recovery mirror for the immutable embedding artifact used by Jarvis Voice in both cloud and local modes.
The normal installation path is Ollama:
ollama pull bigsk1/jarvis-embedding:bf16-v1
Jarvis always calls an Ollama /api/embed endpoint. This Hugging Face repository
is an artifact backup, not a separate runtime embedding provider.
Artifact identity
| Field | Pinned value |
|---|---|
| GGUF file | jarvis-embedding-bf16-v1.gguf |
| GGUF SHA-256 | 0800cbac9c2064dde519420e75e512a83cb360de3ad5df176185dc69652fc515 |
| GGUF size | 621867104 bytes |
| Ollama tag | bigsk1/jarvis-embedding:bf16-v1 |
| Ollama manifest SHA-256 | 85462619ee721b466c5927d109d4cb765861907d5417b9109caebc4e614679f1 |
| Ollama source artifact | embeddinggemma:300m-bf16 |
| Base model | google/embeddinggemma-300m |
| Architecture | Gemma 3 embedding model |
| Parameters | 307.58M |
| Tensor precision | BF16 |
| Output dimensions | 768 |
| Context window | 2048 tokens |
| Pooling | Mean |
The GGUF is byte-for-byte identical to the model layer in the pinned Ollama manifest. Representative query, document, and similarity inputs produced bit-for-bit identical normalized vectors between the upstream Ollama artifact and the versioned Jarvis tag before publication.
Modification notice
Google's upstream EmbeddingGemma model was converted to GGUF BF16 by Ollama.
The Jarvis project did not train, fine-tune, or alter the model tensors from
Ollama's embeddinggemma:300m-bf16 artifact. Jarvis assigned a versioned Ollama
alias and published this recovery mirror and its provenance metadata.
Prompt contract
Jarvis applies Google's asymmetric EmbeddingGemma prompt formats before calling Ollama:
Query: task: search result | query: {text}
Document: title: {title-or-none} | text: {text}
Similarity: task: sentence similarity | query: {text}
Persisted Jarvis embeddings also fingerprint the model tag, Ollama manifest digest, dimensions, prompt profile, and prompt role. A mismatch fails closed so vectors from incompatible artifacts cannot be mixed.
Verify the download
sha256sum -c SHA256SUMS
The expected model-layer digest is 0800cbac9c2064dde519420e75e512a83cb360de3ad5df176185dc69652fc515.
Rebuild into Ollama when the registry is unavailable
Download the repository after acknowledging the Gemma terms, then create a new
local recovery tag from the included Modelfile:
hf download bigsk1/jarvis-embedding-GGUF --local-dir jarvis-embedding-GGUF
cd jarvis-embedding-GGUF
sha256sum -c SHA256SUMS
ollama create jarvis-embedding-recovered:bf16-v1 -f Modelfile
Important: this is a weight-level rebuild, not a drop-in restoration of the pinned Ollama manifest. In a live recovery test, Ollama 0.32.13 reserialized the GGUF and added a template layer. Query, document, and similarity vectors remained bit-for-bit identical at 768 dimensions, but the rebuilt manifest digest differed from the pinned digest.
To adopt a rebuilt artifact, publish it under a new immutable tag, update the
configured model fingerprint, and rebuild every persisted Jarvis embedding
namespace. Never make the existing bf16-v1 identity accept a different digest.
The exact original manifest and all of its source layer bytes are archived here
for a registry-level restoration workflow; do not modify Ollama's private model
store directly.
Provenance files
ollama/manifest-bf16-v1.jsonis the exact registry manifest body.ollama/config.jsonandollama/params.jsonpreserve the exact small Ollama layers referenced by that manifest.LICENSEpreserves the Gemma Terms bundled in the Ollama artifact.Modelfileis a convenience rebuild recipe; it does not promise the same Ollama manifest digest after import.NOTICEcontains Google's required redistribution notice.ARTIFACT.jsonprovides machine-readable identity metadata.
License
This artifact is provided under and subject to the
Gemma Terms of Use and the incorporated
Gemma Prohibited Use Policy.
See LICENSE and NOTICE in this repository.