# Embeddings API

{% hint style="info" %}
The embeddings API is available from Pico 1.1.12 and newer
{% endhint %}

### Supported models

Unlike LLMs, embedding models are downloaded and loaded into memory dynamically when required. Pre-downloading and persistent caching of embedding models is neither necessary nor supported in Pico AI Server. Due to their compact size, embedding models can be efficiently retrieved on demand. 

By default, embedding models are stored in `<base-directory>/Embeddings` as specified in the **Models** section of **Settings**.

Embedding models leverage the CoreML-based [Swift-Embeddings](https://github.com/apple/swift-embeddings) library. For more implementation details, refer to the project’s GitHub repository.

<table><thead><tr><th width="465">Name</th><th>Architecture</th></tr></thead><tbody><tr><td><code>all-MiniLM-L6-v2</code></td><td>BERT </td></tr><tr><td><code>msmarco-bert-base-dot-v5</code></td><td>BERT</td></tr><tr><td><code>LaBSE</code></td><td>BERT</td></tr><tr><td><code>gte-base</code></td><td>BERT</td></tr><tr><td><code>bert-base-uncased</code></td><td>BERT</td></tr><tr><td><code>mxbai-embed-large-v1</code></td><td>BERT</td></tr><tr><td><code>roberta-base</code></td><td>RoBERTa </td></tr><tr><td><code>xlm-roberta-base</code></td><td>XLM-RoBERTa</td></tr><tr><td><code>multilingual-e5-large</code></td><td>XLM-RoBERTa</td></tr><tr><td><code>multilingual-e5-small</code></td><td>XLM-RoBERTa</td></tr><tr><td><code>paraphrase-multilingual-mpnet-base-v2</code></td><td>XLM-RoBERTa</td></tr><tr><td><code>xlm-roberta-base-multilingual-en-ar-fr-de-es-tr-it</code></td><td>XLM-RoBERTa</td></tr><tr><td><code>clip-vit-base-patch16</code></td><td>CLIP</td></tr><tr><td><code>clip-vit-base-patch32</code></td><td>CLIP</td></tr><tr><td><code>clip-vit-large-patch14</code></td><td>CLIP</td></tr><tr><td><code>glove-twitter-25</code></td><td>Word2Vec</td></tr><tr><td><code>glove-twitter-50</code></td><td>Word2Vec</td></tr><tr><td><code>glove-twitter-100</code></td><td>Word2Vec</td></tr><tr><td><code>glove-twitter-200</code></td><td>Word2Vec</td></tr><tr><td><code>potion-base-2M</code></td><td>Model2Vec</td></tr><tr><td><code>potion-base-4M</code></td><td>Model2Vec</td></tr><tr><td><code>potion-base-8M</code></td><td>Model2Vec</td></tr><tr><td><code>potion-retrieval-32M</code></td><td>Model2Vec</td></tr><tr><td><code>potion-base-32M</code></td><td>Model2Vec</td></tr><tr><td><code>M2V_base_output</code></td><td>Model2Vec</td></tr><tr><td><code>M2V_base_output</code></td><td>Static</td></tr><tr><td><code>static-similarity-mrl-multilingual-v1</code></td><td>Static</td></tr></tbody></table>

### OpenAI-compatible endpoint

#### Request

<mark style="color:green;">`POST`</mark> `/v1/embeddings`

| Name              | Type                       | Description                                           |
| ----------------- | -------------------------- | ----------------------------------------------------- |
| `model`           | String                     | Name of the embeddings model, e.g. `all-MiniLM-L6-v2` |
| `input`           | String or array of strings | String or strings to embed                            |
| `encoding_format` | Optional string            | This property is ignored                              |
| `user`            | Optional string            | This property is ignored                              |

#### Response

| Name     | Type                       | Description                                           |
| -------- | -------------------------- | ----------------------------------------------------- |
| `model`  | String                     | Name of the embeddings model, e.g. `all-MiniLM-L6-v2` |
| `object` | String                     | Always `list`                                         |
| `data`   | Array of embedding objects | See embedding objects                                 |

| Name        | Type            | Description            |
| ----------- | --------------- | ---------------------- |
| `index`     | Integer         | Index of the embedding |
| `object`    | String          | Always `embedding`     |
| `embedding` | Array of floats | Embedding              |

### Ollama-compatible endpoint

#### Request

<mark style="color:green;">`POST`</mark> `/api/embed`

| Name    | Type                       | Description                                           |
| ------- | -------------------------- | ----------------------------------------------------- |
| `model` | String                     | Name of the embeddings model, e.g. `all-MiniLM-L6-v2` |
| `input` | String or array of strings | String or strings to embed                            |

#### Response

| Name                | Type                     | Description                               |
| ------------------- | ------------------------ | ----------------------------------------- |
| `model`             | String                   | The model used to generate the embeddings |
| `embeddings`        | Array of array of floats | The generated embeddings                  |
| `total_duration`    | Integer                  | Always `nil`                              |
| `load_duration`     | Integer                  | Always `nil`                              |
| `prompt_eval_count` | Integer                  | Always `nil`                              |