Description #

Text embedding model based on BGE-M3, supports generation of sparse, dense, and token-level vectors

Key features #

• Multi-level embedding output:
  • Sparse vector: term weight representation, suitable for keyword retrieval
  • Dense vector: 1024-dimensional semantic representation, suitable for semantic similarity calculation
  • Token vector: fine-grained contextual representation
• Hardware acceleration: supports FP16 quantization and GPU parallel computation

Operator parameters #

Input #

Output #

Processed array containing the following fields:

• dense_embedding: dense embedding vector
• sparse_embedding: sparse embedding vector
• token_embedding: optional token-level embedding vector

Parameters #

If a parameter does not have a default value, it is required

Examples #

The following code demonstrates how to use daft to run the operator to compute text dense embedding, sparse embedding, and token embedding based on the bge-m3 model.

from __future__ import annotations

import os

import daft
from daft import col
from daft.las.functions.text.embedding.bge_sparse_dense_embedding import BgeSparseDenseEmbedding
from daft.las.functions.udf import las_udf

if __name__ == "__main__":

    if os.getenv("DAFT_RUNNER", "native") == "ray":
        import logging

        import ray

        def configure_logging():
            logging.basicConfig(
                level=logging.INFO,
                format="%(asctime)s - %(name)s - %(levelname)s - %(message)s",
                datefmt="%Y-%m-%d %H:%M:%S.%s".format(),
            )
            logging.getLogger("tracing.span").setLevel(logging.WARNING)
            logging.getLogger("daft_io.stats").setLevel(logging.WARNING)
            logging.getLogger("DaftStatisticsManager").setLevel(logging.WARNING)
            logging.getLogger("DaftFlotillaScheduler").setLevel(logging.WARNING)
            logging.getLogger("DaftFlotillaDispatcher").setLevel(logging.WARNING)

        ray.init(dashboard_host="0.0.0.0", runtime_env={"worker_process_setup_hook": configure_logging})
        daft.set_runner_ray()
    daft.set_execution_config(actor_udf_ready_timeout=600)
    daft.set_execution_config(min_cpu_per_task=0)

    samples = {"text": ["Hello World!", None]}
    is_output_token_vec = True
    dtype = "float16"
    batch_size = 512
    model_path = os.getenv("MODEL_PATH", "/opt/las/models")
    model_name = "BAAI/bge-m3"
    rank = 0

    ds = daft.from_pydict(samples)
    ds = ds.with_column(
        "embeddings",
        las_udf(
            BgeSparseDenseEmbedding,
            construct_args={
                "is_output_token_vec": is_output_token_vec,
                "dtype": dtype,
                "batch_size": batch_size,
                "model_path": model_path,
                "model_name": model_name,
                "rank": rank,
            },
            num_gpus=1,
            batch_size=1,
            concurrency=1,
        )(col("text")),
    )

    ds.show()
    # ╭──────────────┬───────────────────────────────────────────────────────────────────────────────────────────────╮
    # │ text         ┆ embeddings                                                                                    │
    # │ ---          ┆ ---                                                                                           │
    # │ Utf8         ┆ Struct[dense_embedding: List[Float32], sparse_embedding: Map[Utf8: Float32], token_embedding: │
    # │              ┆ List[List[Float32]]]                                                                          │
    # ╞══════════════╪═══════════════════════════════════════════════════════════════════════════════════════════════╡
    # │ Hello World! ┆ {dense_embedding: [-0.0420532…                                                                │
    # ├╌╌╌╌╌╌╌╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌┤
    # │ None         ┆ {dense_embedding: None,                                                                       │
    # │              ┆ spars…                                                                                        │
    # ╰──────────────┴───────────────────────────────────────────────────────────────────────────────────────────────╯