Weighted retrieval for in-context tabular model

A method comprising:

• converting a target record into a target embedding and a plurality of rows into a plurality of embeddings in an embedding space;
• identifying a subset of records that will fit into a limited-size memory of an in-context learning model and are most closely related to the target record;
• updating the in-context learning model by storing the subset of records within the limited-size memory;
• executing the updated in-context learning model a plurality of times on the target record to generate a plurality of tokens, respectively;
• and
• outputting the plurality of tokens as a predictive result.

A computer-readable storage medium comprising instructions which when executed by a computer cause a processor to perform:

• converting a target record into a target embedding and a plurality of rows into a plurality of embeddings in an embedding space;
• identifying a subset of records that will fit into a limited-size memory of an in-context learning model and are most closely related to the target record;
• updating the in-context learning model by storing the subset of records within the limited-size memory;
• executing the updated in-context learning model a plurality of times on the target record to generate a plurality of tokens, respectively;
• and
• outputting the plurality of tokens as a predictive result.