compute-ai-embeddings

This agent uses the configured AI model’s embedding feature to transform a string of text into a vector embedding. At present, it is assumed that only one AI mode will be set in configuration.yaml. This agent will discover its type (ie OpenAI, Hugging Face, Vertex) and use the corresponding library to generate the embedding. It is up to the developer to match the correct embedding model with the configured AI model.

JSON and String inputs

This agent currently only accepts JSON-formatted inputs.

Either ensure the input is JSON, or put the document-to-json agent before the compute-ai-embeddings agent in your pipeline:

pipeline:
  - name: "convert-to-json"
    type: "document-to-json"
    input: "input-topic"
    configuration:
      text-field: "question"
  - name: "compute-embeddings"
    id: "step1"
    type: "compute-ai-embeddings"
    input: "input-topic"
    output: "output-topic"
    configuration:
      model: "${secrets.open-ai.embeddings-model}" # This needs to match the name of the model deployment, not the base model
      embeddings-field: "value.embeddings"
      text: "{{ value.name }} {{ value.description }}"
      batch-size: 10
      # this is in milliseconds. It is important to take this value into consideration when using this agent in the chat response pipeline
      # in fact this value impacts the latency of the response
      # for latency sensitive applications, consider to set batch-size to 1 or flush-interval to 0
      flush-interval: 500

Using Open AI

Set up the OpenAI LLM configuration. Add the compute-ai-embeddings agent:

- name: "compute-embeddings"
  type: "compute-ai-embeddings"
  input: "input-topic" # optional
  output: "output-topic" # optional
  configuration:
    model: "text-embedding-ada-002"
    embeddings-field: "value.embeddings"
    text: "{{ value }}"

Using Google Vertex AI

Set up the Vertex LLM configuration. Add the compute-ai-embeddings agent:

- name: "compute-embeddings"
  type: "compute-ai-embeddings"
  input: "input-topic" # optional
  output: "output-topic" # optional
  configuration:
    model: "textembedding-gecko"
    embeddings-field: "value.embeddings"
    text: "{{ value }}"

Using Ollama

Refer to the ollama documentation to find a list of models.

Setup the Ollama configuration.

Add the compute-ai-embeddings agent:

- name: "compute-embeddings"
  type: "compute-ai-embeddings"
  input: "input-topic" # optional
  output: "output-topic" # optional
  configuration:
    model: "llama2"
    embeddings-field: "value.embeddings"
    text: "{{ value }}"

Ollama models may compute embeddings but they are currently not as good as models provided by OpenAI or Huggingface. Ollama will provide models specifically for embeddings in the future.

Using Amazon Bedrock

Set up the Amazon Bedrock LLM configuration. Add the compute-ai-embeddings agent:

- name: "compute-embeddings"
  type: "compute-ai-embeddings"
  input: "input-topic" # optional
  output: "output-topic" # optional
  configuration:
    model: "amazon.titan-embed-text-v1"
    embeddings-field: "value.embeddings"
    text: "{{ value }}"

Using Huggingface

Set up the Huggingface resource configuration. Add the compute-ai-embeddings agent:

  - name: "compute-embeddings"
    id: "step1"
    type: "compute-ai-embeddings"
    input: "input-topic"
    output: "output-topic"
    configuration:
      model: "${secrets.open-ai.embeddings-model}" # This needs to match the name of the model deployment, not the base model
      embeddings-field: "value.embeddings"
      text: "{{ value.name }} {{ value.description }}"
      batch-size: 10
      # this is in milliseconds. It is important to take this value into consideration when using this agent in the chat response pipeline
      # in fact this value impacts the latency of the response
      # for latency sensitive applications, consider to set batch-size to 1 or flush-interval to 0
      flush-interval: 500

Set HUGGING_FACE_PROVIDER=api and provide your Huggingface key and embeddings model to use the HF inference API:

export HUGGING_FACE_PROVIDER=api
export HUGGING_FACE_ACCESS_KEY=your_access_key
export HUGGING_FACE_EMBEDDINGS_MODEL=multilingual-e5-small

To compute text embeddings with a local model instead of calling the Huggingface API, set HUGGING_FACE_PROVIDER=local and set your embeddings model.

HUGGING_FACE_PROVIDER=local
HUGGING_FACE_EMBEDDINGS_MODEL=multilingual-e5-small
HUGGING_FACE_EMBEDDINGS_MODEL_URL=djl://ai.djl.huggingface.pytorch/intfloat/multilingual-e5-small

The above example will use the multilingual-e5-small Huggingface model locally via the Deep Java Library.

Automatically computing the embeddings over a list of inputs

It is possible to perform the same computation over a list of inputs - for example, a list of questions. You can take the Flare pattern as an example.

In the example below we use the 'loop-over' capability to compute the embeddings for each document in the list of documents to retrieve.

  - name: "compute-embeddings"
    type: "compute-ai-embeddings"
    configuration:
      loop-over: "value.documents_to_retrieve"
      model: "${secrets.open-ai.embeddings-model}"
      embeddings-field: "record.embeddings"
      text: "{{ record.text }}"

When you use "loop-over", the agent executes for each element in a list instead of operating on the whole message. Use "record.xxx" to refer to the current element in the list.

The snippet above computes the embeddings for each element in the list "documents_to_retrieve". The list is expected to be a struct like this:

{
  "documents_to_retrieve": [
      {
        "text": "the text of the first document"
      },
      {
        "text": "the text of the second document"
      }
    ]
}

After running the agent the contents of the list are:

{
  "documents_to_retrieve": [
      {
        "text": "the text of the first document",
        "embeddings": [1,2,3,4,5]
       },
       {
        "text": "the text of the second document",
        "embeddings": [6,7,8,9,10]
       }
    ]
}

Topics

Input

• Structured and unstructured text ?

• Implicit topic ?

• Templating ?

Output

• Structured text ?

• Implicit topic ?

Configuration

Check out the full configuration properties in the API Reference page.