defenx-nlp

Lightweight semantic NLP building blocks for Python.

defenx-nlp gives you one interface for text embeddings, semantic retrieval, prototype-based inference, and simple end-to-end NLP pipelines. It is designed for developers who want production-friendly primitives without wiring the same boilerplate in every project.

What It Does

The package currently covers four layers:

• SemanticEncoder: a backend-driven embedding facade for local transformer models.
• SemanticSearchEngine: semantic indexing and retrieval over embedded documents.
• PrototypeInferenceEngine: lightweight embedding-based classification and scoring.
• NLPipeline: preprocessing -> encode -> infer orchestration with structured output.

This makes the project useful for:

• support ticket routing
• internal knowledge search
• FAQ and help center retrieval
• anomaly or incident scoring
• semantic deduplication and clustering
• retrieval-augmented backends

Who Uses It

This is primarily a developer library, not a direct end-user application.

Typical users are:

• Python backend developers
• ML engineers building semantic features
• support tooling teams
• security/SOC teams experimenting with event similarity
• teams building internal search or classification workflows

End users would normally interact with it indirectly inside:

• a FastAPI or Flask service
• a chatbot or RAG system
• a support desk platform
• an admin dashboard
• a data processing or analytics job

Architecture

Installation

Standard install

pip install defenx-nlp

This installs the package and its core dependencies for a normal CPU workflow.

CUDA install

If you want a CUDA-enabled PyTorch build, reinstall torch with the matching wheel after installing the package:

pip install defenx-nlp
pip install --upgrade torch --index-url https://download.pytorch.org/whl/cu128

Development install

git clone https://github.com/defenx-sec/defenx-nlp.git
cd defenx-nlp
pip install -e ".[dev]"

Quick Start

1. Encode text

from defenx_nlp import SemanticEncoder

enc = SemanticEncoder()

embedding = enc.encode("Neural networks are useful for semantic search.")
print(embedding.shape)  # (384,)

embeddings = enc.encode_batch(["hello", "goodbye", "help me"])
print(embeddings.shape)  # (3, 384)

2. Semantic retrieval

from defenx_nlp import SemanticEncoder, SemanticSearchEngine

enc = SemanticEncoder()
search = SemanticSearchEngine(enc)

search.index(
    [
        "Reset your password",
        "Check your latest invoice",
        "Troubleshoot login issues",
    ]
)

results = search.search("I cannot sign in to my account", top_k=2)
for match in results:
    print(match.rank, round(match.score, 3), match.document.text)

3. Prototype-based classification

from defenx_nlp import SemanticEncoder, PrototypeInferenceEngine

enc = SemanticEncoder()
engine = PrototypeInferenceEngine.from_texts(
    enc,
    {
        "support": ["reset password", "cannot log in", "account help"],
        "billing": ["charged twice", "refund request", "invoice issue"],
    },
)

prediction = engine.infer(enc.encode("please help me reset my login"))
print(prediction.label)
print(prediction.score)

4. Run a simple pipeline

from defenx_nlp import (
    NLPipeline,
    PreprocessingConfig,
    PrototypeInferenceEngine,
    SemanticEncoder,
)

enc = SemanticEncoder()
inference = PrototypeInferenceEngine.from_texts(
    enc,
    {
        "support": ["reset password", "login problem"],
        "billing": ["refund request", "invoice problem"],
    },
)

pipeline = NLPipeline(
    enc,
    inference_engine=inference,
    preprocessing_config=PreprocessingConfig(lowercase=True),
)

result = pipeline.run("HELP! I cannot access my account.")
print(result.processed_text)
print(result.prediction.label)

Why Use This Instead Of Raw sentence-transformers?

You can absolutely use sentence-transformers directly. This project becomes helpful when you want a cleaner application-facing layer around embeddings.

Problem	Raw sentence-transformers	defenx-nlp
Device selection	You handle CPU/CUDA/MPS decisions yourself	get_device() is built in
Service-friendly facade	Model code leaks into app logic	SemanticEncoder keeps a stable interface
Retrieval layer	You wire indexing and ranking yourself	SemanticSearchEngine is ready to use
Simple classifier	You build your own prototype scoring	PrototypeInferenceEngine is included
End-to-end flow	You orchestrate each step manually	NLPipeline returns structured results
Output consistency	Mix of tensors/arrays depending on flags	Returns float32 NumPy arrays

API Summary

Symbol	Description
SemanticEncoder	Main embedding facade
SemanticSearchEngine	Document indexing and semantic retrieval
NumpyVectorIndex	NumPy-based cosine similarity index
FaissVectorIndex	Optional FAISS-backed vector index
PrototypeInferenceEngine	Prototype-based classifier/scoring engine
NLPipeline	Preprocess -> encode -> infer pipeline
EncoderConfig	Backend configuration object
PreprocessingConfig	Cleaning/truncation config for the pipeline
DocumentRecord	Structured retrieval document
SearchResult	Ranked retrieval result
Prediction	Structured inference output
PipelineResult	Structured pipeline output
clean_text, batch_clean, truncate	Preprocessing helpers
cosine_similarity, batch_cosine_similarity	Similarity helpers
normalize_embedding, normalize_batch	L2 normalization helpers

Full API docs: docs/api_reference.md

Backends

The default backend is sentence-transformers.

The package also exports backend contracts for future extension:

• SentenceTransformerBackend: implemented and production-usable
• OnnxEncoderBackend: interface stub, not implemented yet
• APIEncoderBackend: interface stub, not implemented yet

If you expose ONNX or remote API backends publicly, label them as experimental until they perform real inference.

Examples

python examples/basic_usage.py
python examples/batch_encoding.py
python examples/v2_pipeline.py

Testing

pytest tests -v

The test suite contains both:

• pure local unit tests for retrieval, inference, and pipeline logic
• integration-style encoder tests that require the default model to be locally available or downloadable

If the environment cannot reach Hugging Face and the model is not cached, the integration tests skip instead of failing the entire local test run.

Project Structure

defenx-nlp/
|-- defenx_nlp/
|   |-- __init__.py
|   |-- backends.py
|   |-- device.py
|   |-- encoder.py
|   |-- inference.py
|   |-- interfaces.py
|   |-- pipeline.py
|   |-- preprocessing.py
|   |-- retrieval.py
|   |-- schemas.py
|   `-- utils.py
|-- docs/
|   |-- api_reference.md
|   `-- architecture.png
|-- examples/
|   |-- basic_usage.py
|   |-- batch_encoding.py
|   `-- v2_pipeline.py
|-- tests/
|   |-- test_encoder.py
|   `-- test_v2.py
|-- pyproject.toml
`-- README.md

Roadmap

Good next milestones for the project:

• implement the ONNX backend
• implement a real API embedding backend
• add persistence helpers for vector indexes
• add FastAPI service examples
• expand benchmark coverage for CPU vs CUDA vs FAISS
• publish hosted documentation

License

MIT. See LICENSE.