This example shows how to use precomputed binary input columns with ONNXEmbeddings.
The model bge-small-en-v1.5 expects fixed-length input arrays of 512 float32 values. These inputs are named input_ids and attention_mask. When you provide a txt column as input, the function generates these arrays automatically.
If your table already contains tokenized inputs, you can bypass tokenization. Use the input_ids and attention_mask BLOB columns directly. This approach uses precomputed binary inputs and does not require a tokenizer.
The input table amazon_reviews_25_binary contains precomputed input tensors:
- input_ids (BLOB): Array of little-endian INT64 values
- attention_mask (BLOB): Array of little-endian INT64 values
select
*
from mldb.ONNXEmbeddings(
on (SELECT CAST(rev_id AS VARCHAR(21)) as rev_id, input_ids, attention_mask FROM amazon_reviews_25_binary)
on (select model_id, model from onnx_models where model_id = 'bge-small-en-v1.5') DIMENSION
USING
Accumulate('rev_id')
ModelOutputTensor('sentence_embedding')
OutputFormat('FLOAT32(384)')
) a
;
*** Query completed. 25 rows found. 385 columns returned.
*** Warning: 3705 Request executed in interpretive EVL mode.
*** Total elapsed time was 7 seconds.
*** Warning: Report has more than 100 column(s).
Only the first 100 column(s) will be displayed.
rev_id emb_0 emb_1 emb_2 emb_3 emb_4 --------------------- ---------------------- ---------------------- ---------------------- ---------------------- ---------------------- ... A26GKZPS079GFF -3.33678461611271E-002 1.32660176604986E-002 3.98636870086193E-002 3.13083566725254E-002 -1.01601090282202E-002 ... A27ZH1AQORJ1L -4.07209852710366E-003 3.12303509563208E-002 5.48462718725204E-002 2.47544385492802E-002 -2.51442417502403E-002 ... A5E9TSD20U9PR -2.55403872579336E-002 4.24333736300468E-002 3.86486761271954E-003 2.48651672154665E-002 8.27911961823702E-003 ... A2S166WSCFIFP5 2.85573978908360E-003 -1.43960937857628E-002 -1.67111344635487E-002 -8.66278819739819E-003 -1.05052180588245E-002 ... A3V1MKC2BVWY48 1.03411199524999E-002 3.96031923592091E-002 2.03205142170191E-002 1.71305481344461E-002 7.52637279219925E-004 ... AUTNO7VDY4H4A -1.18440315127373E-002 2.22351606935263E-002 -5.43525954708457E-003 2.37749889492989E-002 -1.12732723355293E-002 ... ...