jinaai/jina-embeddings-v5-omni-nano-classification

readme: a video embeds its frames; drop the stale imageio claim

sync custom_st.py from harness/src/hf_sources/nano_variant

Run the audio feature extractor whenever a request carries audio (#4)

sync custom_st.py from harness/src/hf_sources/nano_variant

sync vllm_llava_eurobert_audio.py from harness/src/hf_sources/nano_variant

audio: size the audio run from the real Whisper frame mask

audio: size the audio run from the real Whisper frame mask

Support attn_implementation="flash_attention_2"

readme: point users without torchcodec/torchvision (Windows, some Colab) to the av-only model.encode("clip.mp4") path

readme: install instructions — add torchcodec for proc(videos=path) (transformers' video processor selects torchcodec by default; av/imageio are not consulted on that path)

docs: media query/document via encode_query/encode_document; nano requires transformers>=5.0 for multimodal

fix(processor): count images/videos from grid_thw so a single PIL.Image works (was TypeError on len(Image)); byte-identical for list inputs

README.md: mirror retrieval paradigm — add 'Document: ' prefix to Quickstart text input, switch SBERT to encode_document, prefix the vLLM prompt, add prefix-requirement note.

config_sentence_transformers.json: mirror retrieval — set prompts={'document': 'Document: '} and default_prompt_name=null so SBERT users use encode_document(...), AutoModel/vLLM users prepend 'Document: ' manually (same paradigm as retrieval, no query side).

Revert config.json to pre-2026-05-17 content (parent 84442a148e)

Revert config_sentence_transformers.json to pre-2026-05-17 content (parent 84442a148e)

Revert vllm_llava_eurobert_audio.py to pre-2026-05-17 content (parent 84442a148e)

Revert modeling_llava_eurobert_audio.py to pre-2026-05-17 content (parent 84442a148e)

sync vllm_llava_eurobert_audio.py: auto-prepend default_text_prefix for text-only inputs (no-op on retrieval)

sync modeling_llava_eurobert_audio.py: auto-prepend default_text_prefix for text-only inputs (no-op on retrieval)

config.json: add default_text_prefix="Document: " so the modeling code and vLLM plugin auto-prepend this on text-only inputs (matching the text-* twin's SBERT-default behavior across all three code paths).

Set default_prompt_name="document" and prompts.document="Document: " to mirror the matching text-* twin; SBERT-default text vectors are now bit-identical to text-{nano,small}-{classification,clustering,text-matching}.

readme: add frontier plot + EIS section

add omni_frontier.png: omni model frontier plot from paper

readme: add logo, ArXiv/Blog links, sibling-size link, broaden tags

fix(processor): expand image+video without placeholder collision

Initial public release