The textembedding-gecko-multilingual@001 model is a state-of-the-art text embedding model developed by Google, designed to convert textual data into numerical vector representations. It captures semantic meanings and relationships within the text, facilitating various natural language processing (NLP) tasks.
The model supports multiple languages, including but not limited to Arabic, Bengali, English, Spanish, French, Hindi, Chinese.
The textembedding-gecko-multilingual@001 model is based on a dense vector representation architecture similar to that used in large language models (LLMs). It employs advanced deep learning techniques to generate embeddings that reflect the semantic context of the input text.
The model was trained using a diverse dataset generated through a two-step process involving LLMs. The initial step involves generating queries and relevant passages, while the second step ranks these passages to create a fine-tuning dataset. This approach ensures a broad coverage of tasks and enhances the model's performance.
The training data comprises a large corpus of unlabeled passages. The diversity of the training data contributes significantly to the model's ability to understand and generate meaningful embeddings.
The model's knowledge is current as of April 2024.
The training data is designed to be diverse, which helps mitigate biases. However, as with any model, ongoing evaluation is essential to identify and address any potential biases that may arise from the training data.
The textembedding-gecko-multilingual@001 model exhibits impressive performance metrics, particularly when evaluated against the Massive Text Embedding Benchmark (MTEB). This benchmark is a comprehensive evaluation suite that encompasses seven categories of tasks across 56 individual datasets, allowing for a robust assessment of the model's capabilities.
The model achieves an average score of 66.31 with 768-dimensional embeddings. This score positions it as a leading contender among text embedding models, outperforming larger models (up to 7 times larger) and those with higher dimensional embeddings (up to 4096 dimensions) while maintaining a compact size of only 1.2 billion parameters.
The model excels in several core NLP tasks, achieving the following average scores:
Remarkably, the model demonstrates strong zero-shot capabilities, particularly when trained solely on the synthetic FRet dataset. This indicates that it can effectively generalize to unseen tasks without prior exposure to specific datasets, outperforming several competitive baselines.
The model is available on the AI/ML API platform as "textembedding-gecko-multilingual@001".
Detailed API Documentation is available on the AI/ML API website, providing comprehensive guidelines for integration.
The development and deployment of the textembedding-gecko-multilingual model adhere to ethical guidelines that emphasize responsible AI usage. Developers are encouraged to consider the implications of embedding models in their applications, particularly concerning data privacy and potential biases.
License Type: The textembedding-gecko-multilingual@001 model is currently not open-sourced, and its usage is subject to specific licensing agreements defined by Google. Users should review the terms of service and privacy policies associated with the model's deployment.
OpenAI-compatible — swap the base URL and it works with your existing SDK.
| Type | Price |
|---|---|
| Input | |
| Output | |
| Model | Input | Output | Context | Best for |
|---|---|---|---|---|
Textembedding-gecko-multilingual@001 This page | ||||
| Semantic search + RAG | ||||
| Semantic search + RAG | ||||
| Semantic search + RAG | ||||
| Semantic search + RAG |
Textembedding-gecko-multilingual@001 is priced at input $0.0325 / 1M.
Yes, Textembedding-gecko-multilingual@001 can stream responses as they are generated.
Yes, Textembedding-gecko-multilingual@001 supports both function calling and structured outputs.
Textembedding-gecko-multilingual@001 was built by Google.
It is accessed via the v1 endpoint.
