Generate Text Embeddings With AWS Bedrock

An embedding model is a special kind of model built on the same transformer architecture as LLMs. It is trained in a way that contrasts inputs such that related ones have more similar vectors than unrelated ones. This is useful for a variety of things including classification and measuring text similarity. I find the text similarity use case interesting because I'm trialing an approach that combs through all my TILs to find the most similar posts for each.

There are a number of production-grade providers that host API access to embedding models. One of those providers is AWS Bedrock.

Here are all the embedding models it offers. Notice it's not just text. There are image, multilingual, and multimodal models as well.

❯ aws bedrock list-foundation-models --by-output-modality EMBEDDING \
  --query 'modelSummaries[].modelId'
[
    "amazon.nova-2-multimodal-embeddings-v1:0",
    "twelvelabs.marengo-embed-3-0-v1:0",
    "twelvelabs.marengo-embed-2-7-v1:0",
    "cohere.embed-v4:0",
    "amazon.titan-embed-g1-text-02",
    "amazon.titan-embed-text-v1:2:8k",
    "amazon.titan-embed-text-v1",
    "amazon.titan-embed-text-v2:0:8k",
    "amazon.titan-embed-text-v2:0",
    "amazon.titan-embed-image-v1:0",
    "amazon.titan-embed-image-v1",
    "cohere.embed-english-v3:0:512",
    "cohere.embed-english-v3",
    "cohere.embed-multilingual-v3:0:512",
    "cohere.embed-multilingual-v3"
]

Here I run one of my TIL posts through the amazon.titan-embed-text-v2:0 model.

❯ jq -n --rawfile body ~/dev/jbranchaud/til/llm/count-number-of-tokens-in-a-file.md \
    '{inputText: ("llm: Count Number Of Tokens In A File\n\n" + $body)[:8000],
      dimensions: 1024,
      normalize: true}' > body.json

❯ aws bedrock-runtime invoke-model \
  --region us-east-1 \
  --model-id amazon.titan-embed-text-v2:0 \
  --content-type application/json --accept application/json \
  --body fileb://body.json \
  out.json
{
    "contentType": "application/json"
}

I can then use jq to inspect some properties of the output file:

❯ jq '.embedding | length' out.json
1024

❯ jq '.inputTextTokenCount' out.json
326

It doesn't matter whether I passed it a few tokens or hundreds of tokens, it is going to produce an embedding that is 1024 dimensions.

Here is a peek at what is in the file:

❯ jq '.' out.json | head -n 8
{
  "embedding": [
    -0.039784759283065796,
    -0.014103486202657223,
    0.0918450877070427,
    0.024694688618183136,
    -0.007761063985526562,
    -0.013740465976297855,

Where this gets interesting is when I store embeddings like this for a bunch of content and then use that to do cosine similarity searches. Note that because I normalized the embedding, the cosine similarity is just the dot product of two embeddings.

2.9 KiB Raw Blame History Unescape Escape

Generate Text Embeddings With AWS Bedrock

2.9 KiB

Raw Blame History