Embedding (beta)

Create an embedding for a string column using an embedding model.
Supports Sentence transformers and Encoder/Decoder NLP transformers from Hugging Face.
To enable set enrichment parameter toembedding.
For each embedding enrichment, if you want to monitor the embedding vector on fiddler you MUST create a corresponding TextEmbedding using the enrichment’s output column.

Requirements:

Access to Huggingface inference endpoint - https://api-inference.huggingface.co
Huggingface API token

Supported Models:

model_name	size	Type	pooling_method	Notes
BAAI/bge-small-en-v1.5	small	Sentence Transformer
sentence-transformers/all-MiniLM-L6-v2	med	Sentence Transformer
thenlper/gte-base	med	Sentence Transformer		*(default)*
gpt2	med	Encoder NLP Transformer	last_token
distilgpt2	small	Encoder NLP Transformer	last_token
EleuteherAI/gpt-neo-125m	med	Encoder NLP Transformer	last_token
google/bert_uncased_L-4_H-256_A-4	small	Decoder NLP Transformer	first_token	Smallest Bert
bert-base-cased	med	Decoder NLP Transformer	first_token
distilroberta-base	med	Decoder NLP Transformer	first_token
xlm-roberta-large	large	Decoder NLP Transformer	first_token	Multilingual
roberta-large	large	Decoder NLP Transformer	first_token

fdl.ModelInfo.from_dataset_info(
    dataset_info=dataset_info,
    display_name='llm_model',
    model_task=fdl.core_objects.ModelTask.LLM,
    custom_features = [
      fdl.Enrichment(
          name='Question Embedding', # name of the enerichment, will be the vector col
          enrichment='embedding', 
          columns=['question'], # only one allowed per embedding enrichment, must be a text column in dataframe
          config={ # optional
            'model_name': ... # default: 'thenlper/gte-base'
            'pooling_method': ... # choose from '{first/last/mean}_token'. Only required if NOT using a sentence transformer
          }
      ),
      fdl.TextEmbedding(
        name='question_cf', # name of the text embedding custom feature
        source_column='question', # source - raw text
        column='Question Embedding', # the name of the vector - outpiut of the embedding enrichment
      ),
    ]
)

The above example will lead to generation of new column

FDL Question Embedding(vector) : embeddings corresponding to string column question

Note

In the context of Hugging Face models, particularly transformer-based models used for generating embeddings, the pooling_method determines how the model processes the output of its layers to produce a single vector representation for input sequences (like sentences or documents). This is crucial when using these models for tasks like sentence or document embedding, where you need a fixed-size vector representation regardless of the input length.