pizza-embed-zh-small-v1

A high-quality Chinese text embedding model optimized for ONNX Runtime inference. Part of the Pizza Engine model family by Infinilabs.

Model Details

Property	Value
Dimensions	512
Max Sequence Length	512 tokens
Similarity	Cosine
Format	ONNX (opset 18)
Size	~91 MB
Language	Chinese (zh)
Base Model	BAAI/bge-small-zh-v1.5

Files

onnx/model.onnx          # ONNX graph (~476 KB)
onnx/model.onnx.data     # External weights (~91 MB)
tokenizer.json            # HuggingFace tokenizer
tokenizer_config.json     # Tokenizer configuration

Quick Start

Python (ONNX Runtime)

import onnxruntime as ort
from tokenizers import Tokenizer
import numpy as np

tokenizer = Tokenizer.from_file("tokenizer.json")
session = ort.InferenceSession("onnx/model.onnx")

# Tokenize
encoding = tokenizer.encode("搜索引擎技术")
input_ids = np.array([encoding.ids], dtype=np.int64)
attention_mask = np.array([encoding.attention_mask], dtype=np.int64)
token_type_ids = np.zeros_like(input_ids)

# Inference
outputs = session.run(None, {
    "input_ids": input_ids,
    "attention_mask": attention_mask,
    "token_type_ids": token_type_ids,
})

# Mean pooling + L2 normalize
hidden = outputs[0]  # [batch, seq, 512]
mask_expanded = attention_mask[:, :, None].astype(np.float32)
pooled = (hidden * mask_expanded).sum(axis=1) / mask_expanded.sum(axis=1)
embedding = pooled / np.linalg.norm(pooled, axis=-1, keepdims=True)
print(embedding.shape)  # (1, 512)

Rust (Pizza Engine CLI)

use pizza_cli::LocalEmbeddingClient;

let mut client = LocalEmbeddingClient::new("infinilabs/pizza-embed-zh-small-v1", 32)?;
let embeddings = client.embed(&["搜索引擎技术", "全文检索"])?;
// embeddings: Vec<Vec<f32>>, each 512-dim, L2-normalized

Rust (ort + tokenizers)

use ort::session::Session;
use tokenizers::Tokenizer;

let session = Session::builder()?.commit_from_file("onnx/model.onnx")?;
let tokenizer = Tokenizer::from_file("tokenizer.json")?;

let encoding = tokenizer.encode("搜索引擎技术", true)?;
// ... run session, mean-pool, L2-normalize

Use Cases

• Semantic search for Chinese documents
• Text similarity and clustering
• RAG (Retrieval-Augmented Generation) pipelines
• Question-answer matching

Performance

Optimized for CPU inference via ONNX Runtime. Suitable for edge deployment and embedded search engines without GPU requirements.

Citation

@misc{pizza-embed-zh-small-v1,
  title={pizza-embed-zh-small-v1: Chinese Text Embedding for Pizza Engine},
  author={Infinilabs},
  year={2026},
  url={https://huggingface.co/infinilabs/pizza-embed-zh-small-v1}
}

License

MIT