Nlp Slm Patterns

NLP / SLM Patterns

When to Load This Skill

Load when working with: local language models, anonymization/PII detection, text classification, NER, Ollama, vLLM, Presidio, spaCy, Hugging Face Transformers, SLM inference pipelines.

Philosophy

Local models run locally during development, then deploy to YC GPU VM for production. The adapter interface is identical in both environments — only the endpoint URL changes via .env.

Anonymization projects deal with real personal data. Default posture: treat all input as sensitive until proven otherwise. Never log raw text containing PII.

Model Serving Options

Ollama Local Setup

import httpx
from src.project_name.core.config import settings


class OllamaAdapter:
    def __init__(self) -> None:
        self._base_url = settings.ollama_base_url
        self._model = settings.ollama_model

    async def generate(self, prompt: str, system: str = "") -> str:
        async with httpx.AsyncClient(timeout=120.0) as client:
            response = await client.post(
                f"{self._base_url}/api/generate",
                json={
                    "model": self._model,
                    "prompt": prompt,
                    "system": system,
                    "stream": False,
                },
            )
            response.raise_for_status()
            return response.json()["response"]

    async def chat(self, messages: list[dict]) -> str:
        async with httpx.AsyncClient(timeout=120.0) as client:
            response = await client.post(
                f"{self._base_url}/api/chat",
                json={"model": self._model, "messages": messages, "stream": False},
            )
            response.raise_for_status()
            return response.json()["message"]["content"]

Required .env:

OLLAMA_BASE_URL=http://localhost:11434
OLLAMA_MODEL=qwen2.5:7b

Docker Compose for Ollama:

services:
  ollama:
    image: ollama/ollama:latest
    ports:
      - "11434:11434"
    volumes:
      - ollama_data:/root/.ollama
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

volumes:
  ollama_data:

vLLM Production Setup

from openai import AsyncOpenAI
from src.project_name.core.config import settings


class VLLMAdapter:
    def __init__(self) -> None:
        self._client = AsyncOpenAI(
            api_key="EMPTY",
            base_url=settings.vllm_base_url,
        )
        self._model = settings.vllm_model_name

    async def generate(self, prompt: str, system: str = "", max_tokens: int = 1024) -> str:
        response = await self._client.chat.completions.create(
            model=self._model,
            messages=[
                {"role": "system", "content": system},
                {"role": "user", "content": prompt},
            ],
            max_tokens=max_tokens,
            temperature=0.1,
        )
        return response.choices[0].message.content or ""

Required .env (production):

VLLM_BASE_URL=http://your-gpu-vm-ip:8000/v1
VLLM_MODEL_NAME=Qwen/Qwen2.5-7B-Instruct

vLLM launch command on GPU VM:

python -m vllm.entrypoints.openai.api_server \
  --model Qwen/Qwen2.5-7B-Instruct \
  --port 8000 \
  --tensor-parallel-size 1 \
  --max-model-len 32768

Anonymization with Presidio

Microsoft Presidio is the standard for PII detection and anonymization. It supports custom recognizers and operators.

from presidio_analyzer import AnalyzerEngine, RecognizerResult
from presidio_anonymizer import AnonymizerEngine
from presidio_anonymizer.entities import OperatorConfig


class AnonymizationAdapter:
    def __init__(self) -> None:
        self._analyzer = AnalyzerEngine()
        self._anonymizer = AnonymizerEngine()

    def analyze(self, text: str, language: str = "ru") -> list[RecognizerResult]:
        return self._analyzer.analyze(text=text, language=language)

    def anonymize(self, text: str, language: str = "ru") -> str:
        results = self.analyze(text, language)
        anonymized = self._anonymizer.anonymize(
            text=text,
            analyzer_results=results,
            operators={
                "PERSON": OperatorConfig("replace", {"new_value": "<PERSON>"}),
                "PHONE_NUMBER": OperatorConfig("replace", {"new_value": "<PHONE>"}),
                "EMAIL_ADDRESS": OperatorConfig("replace", {"new_value": "<EMAIL>"}),
                "LOCATION": OperatorConfig("replace", {"new_value": "<LOCATION>"}),
            },
        )
        return anonymized.text

Required dependencies:

dependencies = [
    "presidio-analyzer>=2.2.0",
    "presidio-anonymizer>=2.2.0",
    "spacy>=3.7.0",
]

Install spaCy Russian model:

uv run python -m spacy download ru_core_news_sm
uv run python -m spacy download en_core_web_sm

Custom Presidio Recognizer

For domain-specific entities (e.g., employee IDs, contract numbers):

from presidio_analyzer import PatternRecognizer, Pattern


class EmployeeIdRecognizer(PatternRecognizer):
    PATTERNS = [Pattern("EMPLOYEE_ID", r"\bEMP-\d{6}\b", 0.85)]
    CONTEXT = ["employee", "id", "сотрудник"]

    def __init__(self) -> None:
        super().__init__(
            supported_entity="EMPLOYEE_ID",
            patterns=self.PATTERNS,
            context=self.CONTEXT,
        )

Register with analyzer:

from presidio_analyzer import AnalyzerEngine

analyzer = AnalyzerEngine()
analyzer.registry.add_recognizer(EmployeeIdRecognizer())

Transformers Inference Pipeline

from pathlib import Path
import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer, pipeline


class TextClassifierAdapter:
    def __init__(self, model_path: str | Path) -> None:
        self._pipeline = pipeline(
            "text-classification",
            model=str(model_path),
            tokenizer=str(model_path),
            device=0 if torch.cuda.is_available() else -1,
            top_k=None,
        )

    def classify(self, text: str) -> list[dict]:
        results = self._pipeline(text, truncation=True, max_length=512)
        return sorted(results[0], key=lambda x: x["score"], reverse=True)

Privacy Rules

• Never log raw user text in production — log only metadata (length, language, entity counts)
• Anonymize before storing any text in database
• Use ANONYMIZED_TEXT column naming convention to signal anonymized content
• Test anonymization with synthetic PII — never use real personal data in tests

Further Resources

• resources/spacy-ner.md — custom NER with spaCy
• resources/model-quantization.md — GGUF/AWQ quantization for local inference