PALCrossEncoder¶

class hana_ml.text.pal_cross_encoder.PALCrossEncoder(model_version=None, max_token_num=None)¶

PALCrossEncoder processes paired inputs, like (query, document) pairs, together to determine their relationship or similarity, outputting their relevance scores. Unlike bi-encoders, which encode inputs separately into vectors, cross-encoders analyze the interaction between the input pair simultaneously, allowing for precise semantic understanding and high-accuracy results, particularly in re-ranking tasks within information retrieval.

Parameters

model_versionstr, optional

Specifies the version of SAP PAL cross-encoder model to use.

Currently the only valid version is 'SAP_CER.20250701'.

Defaults to None (currently uses 'SAP_CER.20250701' by default).

max_token_numint, optional

The maximum number of tokens allowed for the input of the cross-encoder model.

The default value and value range depend on the model.

Attributes

stats_DataFrame: A DataFrame containing additional messages related to the resulting scores.

Methods

predict(data, key, content[, thread_number, ...])

Calculate relevance scores for paired inputs using a cross-encoder model.

Examples

Suppose you have a HANA DataFrame df with ID column 'ID', content columns 'QUERY' and 'DOC'. To compute relevance scores for pairs in 'QUERY' and 'DOC' with the same IDs, we can create a PALCrossEncoder instance and call its predict() method:

>>> from hana_ml.text.pal_cross_encoder import PALCrossEncoder
>>> ce_model = PALCrossEncoder(model_version='SAP_CER.20250701')
>>> result = ce_model.predict(data=df, key='ID', content=['QUERY', 'DOC'], return_table=True)
>>> print(result.collect())

predict(data, key, content, thread_number=None, batch_size=None, max_token_num=None, return_table=False)¶

Calculate relevance scores for paired inputs using a cross-encoder model.

Parameters

data: DataFrame

Input data containing the paired inputs for relevance score calculation.

key: str

Name of the key (ID) column.

content: list/tuple of str

Names of the two columns that contain the pairs for relevance score calculation.

thread_number: int, optional

Specifies the number of HTTP connections that are established simultaneously to the backend cross-encoder service. The valid range is from 1 to 10.

Defaults to 6.

batch_size: int, optional

Specifies the number of pairs grouped into a single request before sending to the cross-encoder service.

The value range is from 1 to 50.

Defaults to 10.

max_token_num: int, optional

The maximum number of tokens allowed for the input of the cross-encoder model.

If max_token_num is None, the value in self.max_token_num will be used.

Defaults to None.

return_tablebool, optional

Specifies whether to return the DataFrame of the result or not.

If set to False, a numpy array of relevance scores shall be returned instead.

Defaults to False.

Returns

DataFrame or numpy.ndarray

DataFrame or numpy.ndarray containing the relevance scores.

When return_table is True, the returned DataFrame includes an EXT_MSG column. This column is currently always NULL and is reserved for possible future additional messages related to the resulting scores.

Inherited Methods from PALBase¶

Besides those methods mentioned above, the PALCrossEncoder class also inherits methods from PALBase class, please refer to PAL Base for more details.