redup-topicmodel 0.1.3

gRPC inference service for multilingual BigARTM topic embeddings.

redup.python.topicmodel

gRPC inference service for a multilingual BigARTM topic model: document embeddings and topic explanations.

Contract: redup.topicmodel.v1 (GetDocumentsEmbedding, GetTopicExplanation). PyPI package: redup-topicmodel.

Model

Weights and tokenizers are published on Hugging Face:

redup-ai/topicmodel-multilingual

Expected artifact layout:

<data>/
  artm/parameters.bin
  artm/p_wt.bin
  tokenizers.json.gz
  config.json

Download locally:

pip install huggingface_hub
python -c "from huggingface_hub import snapshot_download; print(snapshot_download('redup-ai/topicmodel-multilingual'))"

Configuration

config/config.yaml:

service:
  port: "[::]:9878"

TopicModel:
  artifact_root: /data          # HF artifact root (Docker default mount)
  num_processors: 1

Override without editing the file via servicekit env substitution (section___key):

export TopicModel___artifact_root=/path/to/topicmodel-multilingual

Instead of artifact_root, you can set paths explicitly: model_path (artm/ directory) and bpe_path (tokenizers.json.gz or a tokenizers directory) — also overridable as TopicModel___model_path / TopicModel___bpe_path.

Run with Docker

Pull the published image and mount the downloaded artifacts at /data (as in the config):

docker run --rm -p 9878:9878 \
  -v /path/to/topicmodel-multilingual:/data:ro \
  redup4ai/redup.python.topicmodel:0.1.0-3.11-slim

The service listens for gRPC on port 9878.

Call it from Python:

import asyncio
from redup_proto_topicmodel.client import Client
from redup_proto_topicmodel.redup.topicmodel.v1.topicmodel_pb2 import (
    Document,
    DocumentPack,
)

async def main():
    client = Client("grpc://localhost:9878")
    pack = DocumentPack(documents=[
        Document(
            document_id="doc-en",
            tokens=["hello", "world"],
            modalities={"lang": "en"},
        ),
    ])
    response = await client.get_documents_embedding("example", pack)
    print(response["embeddings"])

asyncio.run(main())

Run locally without Docker

python -m venv .venv && source .venv/bin/activate
pip install -r src/requirements.txt
pip install -e src

export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
export TopicModel___artifact_root=/path/to/topicmodel-multilingual
PYTHONPATH=src python -m redup_topicmodel.service config/config.yaml

Tests

PYTHONPATH=src python -m pytest tests -q -m "not smoke"

Smoke test with a real model (requires TOPICMODEL_DATA or TopicModel___artifact_root):

export TOPICMODEL_DATA=/path/to/topicmodel-multilingual
PYTHONPATH=src python -m pytest tests -q --smoke

License

MIT — see LICENSE.

Parts of the inference logic are adapted from text_categorization (Copyright (c) 2020, Machine Intelligence Team), BSD 3-Clause. See NOTICE and third_party/text_categorization/.