embedding-reference.html

---
# Copyright Vespa.ai. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
title: "Embedding Reference"
redirect_from:
- /documentation/reference/embedding-reference.html
---

<p>Reference configuration for <a href="../embedding.html">embedders</a>.</p>

<h2 id="model-config-reference">Model config reference</h2>
<p>
  Embedder models uses the <a href="config-files.html#model">model</a> type configuration.
  The <em>model</em> type configuration accepts attributes <code>model-id</code>, <code>url</code> or <code>path</code>,
  and multiple of these can be specified as a single config value, where one is used depending on the deployment environment:
</p>
  <ul>
    <li>If a <code>model-id</code> is specified and the application is deployed on Vespa Cloud, the <code>model-id</code> is used.</li>
    <li>Otherwise, if a <code>url</code> is specified, it is used</li>
    <li>Otherwise, <code>path</code> is used.</li>
  </ul>
<p>
  When using <code>path</code>, the model files must be supplied in the
  Vespa <a href="../application-packages.html#deploying-remote-models">application package</a>.
</p>

<h2 id="huggingface-embedder">Huggingface Embedder</h2>
<p>
  An embedder using any <a href="https://huggingface.co/docs/tokenizers/index">Huggingface tokenizer</a>,
  including multilingual tokenizers,
  to produce tokens which is then input to a supplied transformer model in ONNX model format.
</p>
<p>
  The Huggingface embedder is configured in <a href="services.html">services.xml</a>,
  within the <code>container</code> tag:
</p>
<pre>{% highlight xml %}
<container id="default" version="1.0">
    <component id="hf-embedder" type="hugging-face-embedder">
        <transformer-model path="my-models/model.onnx"/>
        <tokenizer-model path="my-models/tokenizer.json"/>
        <prepend>
          <query>query:</query>
          <document>passage:</document>
        </prepend>
    </component>
    ...
</container>
{% endhighlight %}</pre>

<h3 id="huggingface-embedder-reference-config">Huggingface embedder reference config</h3>
<p>In addition to <a href="#embedder-onnx-reference-config">embedder ONNX parameters</a>:</p>
<table class="table">
  <thead>
    <tr>
      <th>Name</th>
      <th>Occurrence</th>
      <th>Description</th>
      <th>Type</th>
      <th>Default</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>transformer-model</td>
      <td>One</td>
      <td>Use to point to the transformer ONNX model file</td>
      <td><a href="#model-config-reference">model-type</a></td>
      <td>N/A</td>
    </tr>
    <tr>
      <td>tokenizer-model</td>
      <td>One</td>
      <td>Use to point to the <code>tokenizer.json</code> Huggingface tokenizer configuration file</td>
      <td><a href="#model-config-reference">model-type</a></td>
      <td>N/A</td>
    </tr>
    <tr>
      <td>max-tokens</td>
      <td>One</td>
      <td>The maximum number of tokens accepted by the transformer model</td>
      <td>numeric</td>
      <td>512</td>
    </tr>
    <tr>
      <td>transformer-input-ids</td>
      <td>One</td>
      <td>The name or identifier for the transformer input IDs</td>
      <td>string</td>
      <td>input_ids</td>
    </tr>
    <tr>
      <td>transformer-attention-mask</td>
      <td>One</td>
      <td>The name or identifier for the transformer attention mask</td>
      <td>string</td>
      <td>attention_mask</td>
    </tr>
    <tr>
      <td>transformer-token-type-ids</td>
      <td>One</td>
      <td>The name or identifier for the transformer token type IDs.  If the model does not use <code>token_type_ids</code> use <code>&lt;transformer-token-type-ids/&gt;</code></td>
      <td>string</td>
      <td>token_type_ids</td>
    </tr>
    <tr>
      <td>transformer-output</td>
      <td>One</td>
      <td>The name or identifier for the transformer output</td>
      <td>string</td>
      <td>last_hidden_state</td>
    </tr>
    <tr>
      <td>pooling-strategy</td>
      <td>One</td>
      <td>How the output vectors of the ONNX model is pooled to obtain a single vector representation. Valid values are <code>mean</code> and <code>cls</code></td>
      <td>string</td>
      <td>mean</td>
    </tr>
    <tr>
      <td>normalize</td>
      <td>One</td>
      <td>A boolean indicating whether to normalize the output embedding vector to unit length (length 1). Useful for <code>prenormalized-angular</code>
      <a href="schema-reference.html#distance-metric">distance-metric</a></td>
      <td>boolean</td>
      <td>false</td>
    </tr>
    <tr>
      <td>prepend</td>
      <td>Optional</td>
      <td>Prepend instructions that are prepended to the text input before tokenization and inference. Useful 
        for models that have been trained with specific prompt instructions. The instructions are prepended to the input text.
        <ul>
          <li>Element &lt;query&gt; - Optional query prepend instruction.</li>
          <li>Element &lt;document&gt; - Optional document prepend instruction.</li>
        </ul>
    <pre>{% highlight xml %}
    <prepend>
      <query>query:</query>
      <document>passage:</document>
    </prepend>{% endhighlight %}</pre>
        </td>
      <td>Optional &lt;query&gt; &lt;document&gt; elements.</td>
      <td></td>
    </tr>
  </tbody>
</table>



<h2 id="bert-embedder">Bert embedder</h2>
<p>
  The Bert embedder is configured in <a href="services.html">services.xml</a>,
  within the <code>container</code> tag:
</p>
<pre>{% highlight xml %}
<container version="1.0">
  <component id="myBert" type="bert-embedder">
    <transformer-model path="models/e5-small-v2.onnx"/>
    <tokenizer-vocab url="https://huggingface.co/intfloat/e5-small-v2/raw/main/vocab.txt"/>
  </component>
</container>
{% endhighlight %}</pre>

<h3 id="bert-embedder-reference-config">Bert embedder reference config</h3>
<p>In addition to <a href="#embedder-onnx-reference-config">embedder ONNX parameters</a>:</p>
<table class="table">
  <thead>
    <tr>
      <th>Name</th>
      <th>Occurrence</th>
      <th>Description</th>
      <th>Type</th>
      <th>Default</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>transformer-model</td>
      <td>One</td>
      <td>Use to point to the transformer ONNX model file</td>
      <td><a href="#model-config-reference">model-type</a></td>
      <td>N/A</td>
    </tr>
    <tr>
      <td>tokenizer-vocab</td>
      <td>One</td>
      <td>Use to point to the Huggingface <code>vocab.txt</code> tokenizer file with valid wordpiece tokens. Does not support <code>tokenizer.json</code> format.</td>
      <td><a href="#model-config-reference">model-type</a></td>
      <td>N/A</td>
    </tr>
    <tr>
      <td>max-tokens</td>
      <td>One</td>
      <td>The maximum number of tokens allowed in the input</td>
      <td>integer</td>
      <td>384</td>
    </tr>
    <tr>
      <td>transformer-input-ids</td>
      <td>One</td>
      <td>The name or identifier for the transformer input IDs</td>
      <td>string</td>
      <td>input_ids</td>
    </tr>
    <tr>
      <td>transformer-attention-mask</td>
      <td>One</td>
      <td>The name or identifier for the transformer attention mask</td>
      <td>string</td>
      <td>attention_mask</td>
    </tr>
    <tr>
      <td>transformer-token-type-ids</td>
      <td>One</td>
      <td>The name or identifier for the transformer token type IDs.  If the model does not use <code>token_type_ids</code> use <code>&lt;transformer-token-type-ids/&gt;</code></td>
      <td>string</td>
      <td>token_type_ids</td>
    </tr>
    <tr>
      <td>transformer-output</td>
      <td>One</td>
      <td>The name or identifier for the transformer output</td>
      <td>string</td>
      <td>output_0</td>
    </tr>
    <tr>
      <td>transformer-start-sequence-token</td>
      <td>One</td>
      <td>The start of sequence token</td>
      <td>numeric</td>
      <td>101</td>
    </tr>
    <tr>
      <td>transformer-end-sequence-token</td>
      <td>One</td>
      <td>The start of sequence token</td>
      <td>numeric</td>
      <td>102</td>
    </tr>
    <tr>
      <td>pooling-strategy</td>
      <td>One</td>
      <td>How the output vectors of the ONNX model is pooled to obtain a single vector representation. Valid values are <code>mean</code> and <code>cls</code></td>
      <td>string</td>
      <td>mean</td>
    </tr>
  </tbody>
</table>

<h2 id="colbert-embedder">colbert embedder</h2>
<p>
  The colbert embedder is configured in <a href="services.html">services.xml</a>,
  within the <code>container</code> tag:
</p>
<pre>{% highlight xml %}
<container version="1.0">
  <component id="colbert" type="colbert-embedder">
    <transformer-model path="models/colbertv2.onnx"/>
    <tokenizer-model url="https://huggingface.co/colbert-ir/colbertv2.0/raw/main/tokenizer.json"/>
    <max-query-tokens>32</max-query-tokens>
    <max-document-tokens>256</max-document-tokens>
  </component>
</container>
{% endhighlight %}</pre>
<p>The Vespa colbert implementation works with default configurations for transformer models that use WordPiece tokenization. 
</p>

<h3 id="colbert-embedder-reference-config">colbert embedder reference config</h3>
<p>In addition to <a href="#embedder-onnx-reference-config">embedder ONNX parameters</a>:</p>
<table class="table">
  <thead>
    <tr>
      <th>Name</th>
      <th>Occurrence</th>
      <th>Description</th>
      <th>Type</th>
      <th>Default</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>transformer-model</td>
      <td>One</td>
      <td>Use to point to the transformer ColBERT ONNX model file</td>
      <td><a href="#model-config-reference">model-type</a></td>
      <td>N/A</td>
    </tr>
    <tr>
      <td>tokenizer-model</td>
      <td>One</td>
      <td>Use to point to the <code>tokenizer.json</code> Huggingface tokenizer configuration file</td>
      <td><a href="#model-config-reference">model-type</a></td>
      <td>N/A</td>
    </tr>
    <tr>
      <td>max-tokens</td>
      <td>One</td>
      <td>Max length of token sequence the transformer-model can handle </td>
      <td>numeric</td>
      <td>512</td>
    </tr>
    <tr>
      <td>max-query-tokens</td>
      <td>One</td>
      <td>The maximum number of ColBERT query token embeddings. Queries are padded to this length. Must be lower than max-tokens</td>
      <td>numeric</td>
      <td>32</td>
    </tr>
    <tr>
      <td>max-document-tokens</td>
      <td>One</td>
      <td>The maximum number of ColBERT document token embeddings. Documents are not padded. Must be lower than max-tokens</td>
      <td>numeric</td>
      <td>512</td>
    </tr>
    <tr>
      <td>transformer-input-ids</td>
      <td>One</td>
      <td>The name or identifier for the transformer input IDs</td>
      <td>string</td>
      <td>input_ids</td>
    </tr>
    <tr>
      <td>transformer-attention-mask</td>
      <td>One</td>
      <td>The name or identifier for the transformer attention mask</td>
      <td>string</td>
      <td>attention_mask</td>
    </tr>
    <tr>
      <td>transformer-mask-token</td>
      <td>One</td>
      <td>The mask token id used for ColBERT query padding</td>
      <td>numeric</td>
      <td>103</td>
    </tr>
    <tr>
      <td>transformer-start-sequence-token</td>
      <td>One</td>
      <td>The start of sequence token id</td>
      <td>numeric</td>
      <td>101</td>
    </tr>
    <tr>
      <td>transformer-end-sequence-token</td>
      <td>One</td>
      <td>The end of sequence token id</td>
      <td>numeric</td>
      <td>102</td>
    </tr>
    <tr>
      <td>transformer-pad-token</td>
      <td>One</td>
      <td>The pad sequence token id</td>
      <td>numeric</td>
      <td>0</td>
    </tr>
    <tr>
      <td>query-token-id</td>
      <td>One</td>
      <td>The colbert query token marker id</td>
      <td>numeric</td>
      <td>1</td>
    </tr>
    <tr>
      <td>document-token-id</td>
      <td>One</td>
      <td>The colbert document token marker id</td>
      <td>numeric</td>
      <td>2</td>
    </tr>
    <tr>
      <td>transformer-output</td>
      <td>One</td>
      <td>The name or identifier for the transformer output</td>
      <td>string</td>
      <td>contextual</td>
    </tr> 
  </tbody>
</table>
<p>The Vespa colbert-embedder uses <code>[unused0]</code>token id 1 for <code>query-token-id</code>,  and <code>[unused1]</code>, 
  token id 2 for <code> document-token-id</code>document marker. Document punctuation chars are filtered (not configurable). 
  The following characters are removed <code>!"#$%&'()*+,-./:;&lt;=&gt;?@[\]^_`{|}~</code>.
</p>


<h3 id="splade-embedder-reference-config">splade embedder reference config</h3>
<p>In addition to <a href="#embedder-onnx-reference-config">embedder ONNX parameters</a>:</p>
<table class="table">
  <thead>
    <tr>
      <th>Name</th>
      <th>Occurrence</th>
      <th>Description</th>
      <th>Type</th>
      <th>Default</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>transformer-model</td>
      <td>One</td>
      <td>Use to point to the transformer ONNX model file</td>
      <td><a href="#model-config-reference">model-type</a></td>
      <td>N/A</td>
    </tr>
    <tr>
      <td>tokenizer-model</td>
      <td>One</td>
      <td>Use to point to the <code>tokenizer.json</code> Huggingface tokenizer configuration file</td>
      <td><a href="#model-config-reference">model-type</a></td>
      <td>N/A</td>
    </tr>
    <tr>
      <td>term-score-threshold</td>
      <td>One</td>
      <td>An optional threshold to increase sparseness, tokens/terms with a score lower than this is not retained.</td>
      <td>numeric</td>
      <td>N/A</td>
    </tr>
    <tr>
      <td>max-tokens</td>
      <td>One</td>
      <td>The maximum number of tokens accepted by the transformer model</td>
      <td>numeric</td>
      <td>512</td>
    </tr>
    <tr>
      <td>transformer-input-ids</td>
      <td>One</td>
      <td>The name or identifier for the transformer input IDs</td>
      <td>string</td>
      <td>input_ids</td>
    </tr>
    <tr>
      <td>transformer-attention-mask</td>
      <td>One</td>
      <td>The name or identifier for the transformer attention mask</td>
      <td>string</td>
      <td>attention_mask</td>
    </tr>
    <tr>
      <td>transformer-token-type-ids</td>
      <td>One</td>
      <td>The name or identifier for the transformer token type IDs.  If the model does not use <code>token_type_ids</code> use <code>&lt;transformer-token-type-ids/&gt;</code></td>
      <td>string</td>
      <td>token_type_ids</td>
    </tr>
    <tr>
      <td>transformer-output</td>
      <td>One</td>
      <td>The name or identifier for the transformer output</td>
      <td>string</td>
      <td>logits</td>
    </tr>
  </tbody>
</table>



<h2 id="huggingface-tokenizer-embedder">Huggingface tokenizer embedder</h2>
  <p>
    The Huggingface tokenizer embedder is configured in <a href="services.html">services.xml</a>,
    within the <code>container</code> tag:
  </p>
<pre>{% highlight xml %}
  <container version="1.0">
    <component id="tokenizer" type="hugging-face-tokenizer">
      <model url="https://huggingface.co/bert-base-uncased/raw/main/tokenizer.json"/>
    </component>
  </container>
{% endhighlight %}</pre>

<h3 id="huggingface-tokenizer-reference-config">Huggingface tokenizer reference config</h3>
<table class="table">
  <thead>
    <tr>
      <th>Name</th>
      <th>Occurrence</th>
      <th>Description</th>
      <th>Type</th>
      <th>Default</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>model</td>
      <td>One To Many</td>
      <td>Use to point to the <code>tokenizer.json</code> Huggingface tokenizer configuration file.
        Also supports <code>language</code>, which is only relevant if one wants to tokenize differently based on the document language.
        Use "unknown" for a model to be used for any language (i.e. by default).</td>
      <td><a href="#model-config-reference">model-type</a></td>
      <td>N/A</td>
    </tr>
  </tbody>
</table>


<h2 id="embedder-onnx-reference-config">Embedder ONNX reference config</h2>
<p>
  Vespa uses <a href="https://onnxruntime.ai/">ONNX Runtime</a> to accelerate inference of embedding models.
  These parameters are valid for both <a href="#bert-embedder">Bert embedder</a> and
  <a href="#huggingface-embedder">Huggingface embedder</a>.
</p>
<table class="table">
  <thead>
    <tr>
      <th>Name</th>
      <th>Occurrence</th>
      <th>Description</th>
      <th>Type</th>
      <th>Default</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>onnx-execution-mode</td>
      <td>One</td>
      <td>Low level ONNX execution model. Valid values are <code>parallel</code> or <code>sequential</code>.
         Only relevant for inference on CPU.
         See <a href="https://onnxruntime.ai/docs/performance/tune-performance/threading.html">ONNX runtime documentation</a> on threading.
        </td>
      <td>string</td>
      <td>sequential</td>
    </tr>
    <tr>
      <td>onnx-interop-threads</td>
      <td>One</td>
      <td>Low level ONNX setting.Only relevant for inference on CPU.</td>
      <td>numeric</td>
      <td>1</td>
    </tr>
    <tr>
      <td>onnx-intraop-threads</td>
      <td>One</td>
      <td>Low level ONNX setting. Only relevant for inference on CPU.</td>
      <td>numeric</td>
      <td>4</td>
    </tr>
    <tr>
      <td>onnx-gpu-device</td>
      <td>One</td>
      <td>The GPU device to run the model on.
        See <a href="/en/operations-selfhosted/vespa-gpu-container.html">configuring GPU for Vespa container image</a>. Use <code>-1</code> to not use GPU for the model, even
        if the instance has available GPUs.
      </td>
      <td>numeric</td>
      <td>0</td>
    </tr>
  </tbody>
</table>

<h2 id="sentencepiece-embedder">SentencePiece embedder</h2>
  <p>
    A native Java implementation of <a href="https://github.com/google/sentencepiece">SentencePiece</a>.
    SentencePiece breaks text into chunks independent of spaces,
    which is robust to misspellings and works with CJK languages.
    Prefer the <a href="#huggingface-tokenizer-embedder">Huggingface tokenizer embedder</a> over this
    for better compatibility with Huggingface models.
  </p>
  <p>
    This is suitable to use in conjunction with <a href="../jdisc/container-components.html">custom components</a>,
    or the resulting tensor can be used in <a href="../ranking.html">ranking</a>.
  </p>
  <p>
    To use the
    <a href="https://github.com/vespa-engine/vespa/blob/master/linguistics-components/src/main/java/com/yahoo/language/sentencepiece/SentencePieceEmbedder.java">
    SentencePiece embedder</a>, add it to <a href="services.html">services.xml</a>:
  </p>
  <pre>{% highlight xml %}
  <container version="1.0">
    <component id="mySentencePiece"
             class="com.yahoo.language.sentencepiece.SentencePieceEmbedder"
             bundle="linguistics-components">
      <config name="language.sentencepiece.sentence-piece">;
          <model>
              <item>
                <language>unknown</language>
                <path>model/en.wiki.bpe.vs10000.model</path>
              </item>
          </model>
        </config>
    </component>
  </container>
  {% endhighlight %}</pre>
  <p>
    See the options available for configuring SentencePiece in
    <a href="https://github.com/vespa-engine/vespa/blob/master/linguistics-components/src/main/resources/configdefinitions/language.sentencepiece.sentence-piece.def">
    the full configuration definition</a>.
  </p>


  <h2 id="wordpiece-embedder">WordPiece embedder</h2>
  <p>
    A native Java implementation of
    <a href="https://github.com/google-research/bert#tokenization">WordPiece</a>,
    which is commonly used with BERT models.
    Prefer the <a href="#huggingface-tokenizer-embedder">Huggingface tokenizer embedder</a>
    over this for better compatibility with Huggingface models.
  </p>
  <p>
    This is suitable to use in conjunction with <a href="../jdisc/container-components.html">custom components</a>,
    or the resulting tensor can be used in <a href="../ranking.html">ranking</a>.
  </p>
  <p>
    To use the
    <a href="https://github.com/vespa-engine/vespa/blob/master/linguistics-components/src/main/java/com/yahoo/language/wordpiece/WordPieceEmbedder.java">
    WordPiece embedder</a>,
    add it to <a href="services.html">services.xml</a> within the <code>container</code> tag:
  </p>
  <pre>{% highlight xml %}
  <container version="1.0">
    <component id="myWordPiece">
             class="com.yahoo.language.wordpiece.WordPieceEmbedder"
             bundle="linguistics-components">
      <config name="language.wordpiece.word-piece">
        <model>
          <item>
            <language>unknown</language>
            <path>models/bert-base-uncased-vocab.txt</path>
          </item>
        </model>
      </config>
    </component>
  </container>
  {% endhighlight %}</pre>
  <p>
    See the options available for configuring WordPiece in
    <a href="https://github.com/vespa-engine/vespa/blob/master/linguistics-components/src/main/resources/configdefinitions/language.wordpiece.word-piece.def">
    the full configuration definition</a>.
  </p>
  <p>
    WordPiece is suitable to use in conjunction with custom components,
    or the resulting tensor can be used in <a href="../ranking.html">ranking</a>.
  </p>

<h2 id="using-an-embedder-from-java">Using an embedder from Java</h2>
<p>
  When writing custom Java components (such as <a href="../searcher-development.html">Searchers</a>
  or <a href="../document-processing.html#document-processors">Document processors</a>),
  use embedders you have configured by
  <a href="../jdisc/injecting-components.html">having them injected in the constructor</a>,
  just as any other component:
</p>
<pre>{% highlight java %}
class MyComponent {
  @Inject
  public MyComponent(ComponentRegistry<Embedder> embedders) {
    // embedders contains all the embedders configured in your services.xml
  }
}
{% endhighlight %}</pre>
<p>See a concrete example of using an embedder in a custom searcher in
<a href="https://github.com/vespa-cloud/vespa-documentation-search/blob/main/src/main/java/ai/vespa/cloud/docsearch/LLMSearcher.java">LLMSearcher</a>.
</p>

<h2 id="custom-embedders">Custom Embedders</h2>
<p>Vespa provides a Java interface for defining components which can provide embeddings of text:
<a href="https://github.com/vespa-engine/vespa/blob/master/linguistics/src/main/java/com/yahoo/language/process/Embedder.java">
com.yahoo.language.process.Embedder</a>.</p>

<p>
  To define a custom embedder in an application and make it usable by Vespa
  (see <a href="../embedding.html#embedding-a-query-text">embedding a query text</a>),
  implement this interface and add it as a <a href="../developer-guide.html#developing-components">component</a> to
  <a href="services-container.html">services.xml</a>:
</p>
<pre>{% highlight xml %}
<container version="1.0">
    <component id="myEmbedder"
      class="com.example.MyEmbedder"
      bundle="the name in artifactId in pom.xml">
        <config name='com.example.my-embedder'>
            <model model-id="minilm-l6-v2"/>
            <vocab path="files/vocab.txt"/>
            <myValue>foo</myValue>
        </config>
    </component>
</container>
{% endhighlight %}</pre>