aiSSEMBLE Inference Sumy Module

Sumy library integration for text summarization support in the aiSSEMBLE Inference library.

Supported Algorithms

This module supports multiple summarization algorithms through the sumy library:

Algorithm	Description
TextRank	Graph-based ranking algorithm (default)
LSA	Latent Semantic Analysis
LexRank	Graph-based ranking with cosine similarity

Installation

pip install aissemble-inference-sumy

This will also install the required dependencies:

• aissemble-inference-core
• sumy (text summarization library)
• nltk (natural language processing toolkit)

Usage

With MLServer

Create a model-settings.json:

{
    "name": "sumy",
    "implementation": "aissemble_inference_sumy.SumyRuntime",
    "parameters": {
        "algorithm": "textrank",
        "sentences_count": 3,
        "language": "english"
    }
}

Start MLServer:

mlserver start /path/to/models

With OIP Client

from aissemble_inference_core.client import InferenceClient
from aissemble_inference_core.client.registry import ModuleRegistry

# The module auto-registers via entry points
print(ModuleRegistry.instance().list_available())
# {'runtimes': ['sumy', ...], 'translators': ['sumy', ...], ...}

# Use with InferenceClient
client = InferenceClient(adapter, endpoint)
result = client.summarize("sumy").text("Long article text...").run()
print(result.summary)

Components

SumyRuntime

MLServer-compatible runtime that wraps sumy summarization algorithms:

• Accepts text strings via OIP protocol
• Returns summarized text
• Configurable algorithm, sentence count, and language via parameters

SumyTranslator

Translator for sumy outputs:

• Inherits from DefaultSummarizationTranslator
• Can be extended with sumy-specific optimizations

Configuration Parameters

Parameter	Type	Default	Description
algorithm	string	"textrank"	Summarization algorithm (textrank, lsa, lexrank)
sentences_count	int	3	Number of sentences in the summary
language	string	"english"	Language for text tokenization

Note: The max_length and min_length parameters from SummarizationBuilder are not currently supported by this module. This is planned for a future enhancement.

Entry Points

This module registers the following entry points:

Group	Name	Class
inference.runtimes	sumy	SumyRuntime
inference.translators	sumy	SumyTranslator

Language Support

The sumy library supports multiple languages for tokenization and stemming. Specify the language using the language parameter in model-settings.json. Common languages include:

• english
• spanish
• french
• german
• italian
• portuguese
• russian
• czech
• slovak

Refer to the sumy documentation for a complete list of supported languages.

Testing

This module includes comprehensive tests covering:

Unit Tests

• Module registration and discovery
• Runtime loading with different algorithms
• Direct inference via SumyRuntime
• Translator functionality
• Error handling and validation

Integration Tests (tagged with @integration)

• End-to-end OIP communication via MLServer
• Real HTTP requests using InferenceClient
• Algorithm switching (TextRank, LSA, LexRank)
• Parameter configuration validation

Run tests:

# Run all tests (unit + integration)
mvn clean install

# Run only unit tests (excluding @integration tag)
behave --tags=-integration

The integration tests automatically:

1. Start MLServer with configured models
2. Execute real inference requests via OIP protocol
3. Validate responses and summaries
4. Clean up resources

See tests/features/ for test scenarios and tests/features/environment.py for MLServer test infrastructure.

Example

For a complete working example, see the aissemble-summarization-example in the examples directory:

cd aissemble-inference-examples/aissemble-summarization-example
uv sync --group test
uv run behave

The example demonstrates:

• MLServer configuration for multiple summarization algorithms
• Integration with InferenceClient and HttpOipAdapter
• BDD test patterns for OIP-based applications
• Sample article data for testing