Comment on page
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.
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: "${secrets.open-ai.embeddings-model}" # This needs to match the name of the model deployment, not the base model
embeddings-field: "value.embeddings"
text: "{{ value.name }} {{ value.description }}"
batch-size: 10
# this is in milliseconds. It is important to take this value into consideration when using this agent in the chat response pipeline
# in fact this value impacts the latency of the response
# for latency sensitive applications, consider to set batch-size to 1 or flush-interval to 0
flush-interval: 500
- name: "compute-embeddings"
type: "compute-ai-embeddings"
input: "input-topic" # optional
output: "output-topic" # optional
configuration:
model: "text-embedding-ada-002"
embeddings-field: "value.embeddings"
text: "{{ value }}"
- name: "compute-embeddings"
type: "compute-ai-embeddings"
input: "input-topic" # optional
output: "output-topic" # optional
configuration:
model: "textembedding-gecko"
embeddings-field: "value.embeddings"
text: "{{ value }}"
Add the
compute-ai-embeddings
agent:- name: "compute-embeddings"
type: "compute-ai-embeddings"
input: "input-topic" # optional
output: "output-topic" # optional
configuration:
model: "llama2"
embeddings-field: "value.embeddings"
text: "{{ value }}"
Ollama models may compute embeddings but they are currently not as good as models provided by OpenAI or Huggingface. Ollama will provide models specifically for embeddings in the future.
- name: "compute-embeddings"
type: "compute-ai-embeddings"
input: "input-topic" # optional
output: "output-topic" # optional
configuration:
model: "amazon.titan-embed-text-v1"
embeddings-field: "value.embeddings"
text: "{{ value }}"
- name: "compute-embeddings"
id: "step1"
type: "compute-ai-embeddings"
input: "input-topic"
output: "output-topic"
configuration:
model: "${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.
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]
}
]
}
Last modified 14d ago