Spacy

spaCy Best Practices

spaCy is the backbone of our NLP systems. Adhering to these guidelines ensures our pipelines are performant, reproducible, and easy to maintain.

1. Project Organization & Configuration

Always structure your spaCy projects using spacy project and define all pipeline settings in a declarative YAML config. This is non-negotiable for reproducibility and scalability.

✅ GOOD: Use spacy project and declarative configs. Organize your pipeline logic in a dedicated src/pipeline/ package. Use project.yml to manage workflows and config.cfg for all spaCy pipeline settings.

# src/pipeline/custom_component.py
from spacy.language import Language
from spacy.tokens import Doc

@Language.factory("my_custom_component")
def create_my_component(nlp: Language, name: str):
    return MyCustomComponent(nlp, name)

class MyCustomComponent:
    def __init__(self, nlp: Language, name: str):
        self.nlp = nlp
        self.name = name

    def __call__(self, doc: Doc) -> Doc:
        # Custom logic here
        return doc

# project.yml (simplified)
# ...
workflows:
  train:
    - "python -m spacy train config.cfg --output models/"
  package:
    - "python -m spacy package models/en_core_web_v1.0.0 ./dist --build wheel"

# config.cfg (simplified)
[nlp]
lang = "en"
pipeline = ["tok2vec", "ner", "my_custom_component"]

[components.my_custom_component]
factory = "my_custom_component"

❌ BAD: Ad-hoc scripts and hardcoded parameters. Avoid scattering pipeline logic across multiple scripts or hardcoding model paths and hyperparameters. This makes experiments non-reproducible and deployment fragile.

# bad_script.py
import spacy
# Parameters hardcoded or passed via CLI args, not in a central config
MODEL_PATH = "path/to/my/model"
THRESHOLD = 0.7

nlp = spacy.load(MODEL_PATH)
# ... pipeline components added programmatically ...

2. Type Hinting

Strictly use type hints for all spaCy objects (Language, Doc, Span, Token). This improves code readability, enables static analysis, and reduces runtime errors.

✅ GOOD: Comprehensive type hints. Annotate function arguments and return types with precise spaCy types.

import spacy
from spacy.language import Language
from spacy.tokens import Doc, Span, Token
from typing import List

def process_document(nlp: Language, text: str) -> Doc:
    """Processes text with a spaCy pipeline."""
    return nlp(text)

def extract_entities(doc: Doc) -> List[Span]:
    """Extracts named entities from a processed Doc."""
    return list(doc.ents)

def get_token_lemma(token: Token) -> str:
    """Returns the lemma of a single token."""
    return token.lemma_

❌ BAD: Untyped spaCy code. Omitting type hints makes code harder to understand and refactor.

# bad_types.py
import spacy

def process_document(nlp, text): # What are nlp and text? What does it return?
    return nlp(text)

def extract_entities(doc): # What is 'doc'?
    return list(doc.ents)

3. Performance Considerations

Optimize for speed and memory by disabling unused pipeline components and processing text in batches.

✅ GOOD: Efficient pipeline usage. Use nlp.select_pipes for inference to only run necessary components. Process multiple documents with nlp.pipe.

import spacy
from spacy.language import Language
from spacy.tokens import Doc
from typing import Iterator

def analyze_text_batch(nlp: Language, texts: List[str]) -> Iterator[Doc]:
    """Processes a batch of texts, only running NER."""
    # Temporarily disable unused pipes for performance during inference
    with nlp.select_pipes(enable=["ner"]):
        yield from nlp.pipe(texts, batch_size=50)

# Example usage
nlp_full = spacy.load("en_core_web_sm")
documents = ["Apple is looking at buying U.K. startup.", "Tim Cook is CEO of Apple."]
for doc in analyze_text_batch(nlp_full, documents):
    print(f"Text: {doc.text}, Entities: {[(ent.text, ent.label_) for ent in doc.ents]}")

❌ BAD: Inefficient processing. Loading full pipelines for simple tasks or processing documents one by one in a loop is wasteful.

# bad_performance.py
import spacy

nlp = spacy.load("en_core_web_lg") # Loads all components, even if only tokenization is needed
texts = ["Text 1", "Text 2", "Text 3"]

for text in texts: # Processes one by one, no batching
    doc = nlp(text)
    # ... only use doc.text or doc.tokens, ignoring NER, POS, etc.

4. Virtual Environments & Dependencies

Always use a virtual environment and pin your spacy version in requirements.txt. This prevents dependency conflicts and ensures consistent environments.

✅ GOOD: Pinned dependencies in a virtual environment.

# Terminal
python -m venv .venv
source .venv/bin/activate
pip install -U pip setuptools wheel
pip install -r requirements.txt
python -m spacy download en_core_web_sm

# requirements.txt
spacy==3.8.*
spacy-llm==0.8.* # If using LLM integration
# Other dependencies...

❌ BAD: Global installs or unpinned versions. Installing spacy globally or using spacy>=3.0 can lead to unexpected behavior and "works on my machine" issues.

# requirements.txt (BAD)
spacy # No version specified, will install latest

5. Custom Components & Extension Attributes

Extend spaCy's Doc, Span, and Token objects using custom components and extension attributes. This keeps your custom logic integrated and accessible.

✅ GOOD: Custom components and extension attributes.

# src/pipeline/custom_component.py
from spacy.language import Language
from spacy.tokens import Doc, Span, Token

# Register extension attribute
if not Doc.has_extension("custom_sentiment"):
    Doc.set_extension("custom_sentiment", default=0.0)

@Language.factory("sentiment_analyzer")
def create_sentiment_analyzer(nlp: Language, name: str):
    return SentimentAnalyzer(nlp, name)

class SentimentAnalyzer:
    def __init__(self, nlp: Language, name: str):
        self.nlp = nlp
        self.name = name

    def __call__(self, doc: Doc) -> Doc:
        # Simple sentiment logic
        score = sum(1 for token in doc if token.text == "good") - sum(1 for token in doc if token.text == "bad")
        doc._.custom_sentiment = float(score)
        return doc

# config.cfg
# ...
[nlp]
pipeline = ["tok2vec", "ner", "sentiment_analyzer"]

[components.sentiment_analyzer]
factory = "sentiment_analyzer"

# Usage
nlp = spacy.load("path/to/your/trained_model") # Load model with custom component
doc = nlp("This is a good movie, not bad at all.")
print(f"Sentiment: {doc._.custom_sentiment}") # Access custom attribute

❌ BAD: Storing custom data separately or modifying core attributes. Avoid maintaining parallel data structures or attempting to directly modify Doc attributes that are meant to be read-only.

# bad_extensions.py
import spacy

nlp = spacy.load("en_core_web_sm")
doc = nlp("Some text.")

# ❌ BAD: Storing custom data in a separate dictionary
custom_data = {doc.text: {"sentiment": 0.5}}

# ❌ BAD: Attempting to modify read-only attributes (will raise error)
# doc.user_data = {"my_field": "value"}

6. Testing

Write focused unit tests for custom components and integration tests for pipeline behavior using pytest. Use minimal nlp objects for unit tests.

✅ GOOD: Targeted unit tests with pytest.

# tests/test_custom_component.py
import pytest
import spacy
from spacy.language import Language
from spacy.tokens import Doc

# Assuming custom_component.py is in src/pipeline/
# from src.pipeline.custom_component import create_my_component, MyCustomComponent

# Register factory for testing
@Language.factory("test_my_custom_component")
def create_test_my_component(nlp: Language, name: str):
    return MyCustomComponent(nlp, name)

class MyCustomComponent: # Define a minimal version for testing
    def __init__(self, nlp: Language, name: str):
        self.nlp = nlp
        self.name = name
    def __call__(self, doc: Doc) -> Doc:
        doc.set_extension("processed_by_custom", default=True, force=True)
        return doc

def test_my_custom_component_adds_extension():
    nlp = spacy.blank("en")
    nlp.add_pipe("test_my_custom_component")
    doc = nlp("Hello world.")
    assert doc._.processed_by_custom is True

def test_my_custom_component_returns_doc():
    nlp = spacy.blank("en")
    nlp.add_pipe("test_my_custom_component")
    doc = nlp("Another test.")
    assert isinstance(doc, Doc)

❌ BAD: Untested components or overly broad tests. Avoid relying solely on manual testing or writing integration tests that load full, complex models for every small change.

# tests/bad_test.py
import spacy

def test_full_pipeline_output():
    # ❌ BAD: Loads a large, slow model for a simple test
    nlp = spacy.load("en_core_web_lg")
    doc = nlp("This is a test sentence.")
    assert len(doc.ents) > 0 # Fragile, depends on model behavior

7. LLM Integration

For integrating Large Language Models, leverage the spacy-llm package. It provides a structured, component-based approach that aligns with spaCy's pipeline philosophy.

✅ GOOD: Use spacy-llm for structured LLM integration. Define LLM components in your config.cfg just like any other spaCy component.

# config.cfg (simplified for spacy-llm)
[nlp]
lang = "en"
pipeline = ["llm_ner"]

[components.llm_ner]
factory = "spacy.LLM"
task = {"@llm_tasks": "spacy.NER.v3"}
model = {"@llm_models": "spacy.GPT-3-5.v1"}

❌ BAD: Ad-hoc LLM calls outside the pipeline. Directly calling LLM APIs in your application code bypasses spaCy's structured pipeline, making it harder to manage, test, and swap models.

# bad_llm.py
import openai # or other LLM client
import spacy

def process_with_llm_and_spacy(text: str):
    # ❌ BAD: LLM call is separate from the spaCy pipeline
    response = openai.Completion.create(
        model="text-davinci-003",
        prompt=f"Extract entities from: {text}",
        max_tokens=100
    )
    llm_entities = parse_llm_response(response)

    nlp = spacy.load("en_core_web_sm")
    doc = nlp(text)
    # Now you have two separate sets of entities, difficult to reconcile
    return doc, llm_entities