Models & input types

Ahnlich AI ships with a set of pre-trained models. Think of them in five families — pick the family that matches your data, then pick a size. Every model lists an embedding dim: two models are compatible (usable as index + query in one store) only when their dims match.

Each model name links to its source. The mark is a Hugging Face model card; the mark is a project repository.

📝 Text models

For sentences, paragraphs, FAQs, product descriptions — anything written.

Model	string name	Max input	Embedding dim	Best for
ALL-MiniLM-L6-v2	all-minilm-l6-v2	256 tokens	384	Fast, lightweight semantic similarity (FAQ search, chatbots).
ALL-MiniLM-L12-v2	all-minilm-l12-v2	256 tokens	384	Higher accuracy for nuanced text, a bit more compute.
BGE-Base-EN-v1.5	bge-base-en-v1.5	512 tokens	768	Balanced speed/quality for production.
BGE-Large-EN-v1.5	bge-large-en-v1.5	512 tokens	1024	Highest precision when latency matters less.

🖼️ Image models

For photos, product pictures, and other images.

Model	string name	Max input	Embedding dim	Best for
ResNet-50	resnet-50	224×224 px	2048	Image-to-image similarity and clustering.
CLIP ViT-B/32 (image)	clip-vit-b32-image	224×224 px	512	The image half of text↔image search (pairs with the CLIP text model).

🧑 Face models

Detect and recognize faces in an image. These return one embedding per detected face, plus metadata (bounding box, confidence, optional age/gender).

Model	string name	Max input	Embedding dim	Best for
Buffalo-L	buffalo-l	640×640 px	512	Detection + recognition, with optional age/gender. Non-commercial use only.
SFace + YuNet	sface-yunet	640×640 px	128	Lightweight detection + recognition. Apache 2.0 / MIT — commercially usable.

See model constraints for their input rules.

🔊 Audio models

For sound clips — music, speech, effects.

Model	string name	Max input	Embedding dim	Best for
CLAP (audio)	clap-audio	10 sec	512	Audio similarity; the audio half of text↔audio search.

See audio constraints.

💻 Code models

For source code across many languages.

Model	string name	Max input	Embedding dim	Best for
Jina Code v2	jina-embeddings-v2-base-code	8192 tokens	768	Code search over 30+ languages — see Code search.

Cross-modal pairs

Some models come as a matched pair that share a dimension, so you can index one modality and query with another. Use one as the INDEXMODEL and its partner as the QUERYMODEL.

Text encoder	Media encoder	Shared dim	Enables
clip-vit-b32-text (77 tokens)	clip-vit-b32-image	512	Text-to-image / image-to-text search
clap-text (512 tokens)	clap-audio	512	Text-to-audio / audio-to-text search

# search product photos with words: index images, query with text$CREATESTORE lookbook INDEXMODEL clip-vit-b32-image QUERYMODEL clip-vit-b32-text

Supported input types

Whatever the model, you hand the proxy one of three raw input kinds and it does the embedding:

Input type	What it is	Embedded by
RAW_STRING	Natural text — sentences, paragraphs, code	a text/code model
IMAGE	An image file	an image or face model
AUDIO	An audio clip	an audio model

Quick pick

If you're working with…	Start with
Text, and want speed	all-minilm-l6-v2
Text, and want accuracy	bge-large-en-v1.5
Images only	resnet-50
Searching images by text	CLIP pair (clip-vit-b32-*)
Faces (commercial)	sface-yunet
Audio	CLAP pair (clap-*)
Source code	jina-embeddings-v2-base-code