nerguard 1.0.2

Entropy-gated hybrid NER for privacy-compliant PII detection

NerGuard

Entropy-Gated Hybrid NER for Privacy-Compliant PII Detection

🤗 Model on HuggingFace  ·  📦 PyPI: nerguard  · 

NerGuard is a pre-ingestion privacy layer for RAG pipelines: it detects and redacts PII from text before documents are indexed, keeping sensitive data out of vector databases and LLM context windows. It runs a multilingual mDeBERTa-v3 base model for fast, high-confidence predictions, then selectively routes only uncertain spans to an LLM (OpenAI or local Ollama) for correction, typically less than 3% of tokens. A three-stage regex layer handles structured PII (credit cards, SSNs, IBANs) with deterministic validation. The result is a hybrid pipeline that matches or exceeds larger models on PII recall while remaining GDPR-auditable: every prediction carries its source, confidence score, and routing decision.

Install

pip install nerguard

The NER model (~300 MB) downloads automatically from HuggingFace on first use.

Quick start

from nerguard import Redactor

ng = Redactor(
    model_path=None,        # str  — local path or HuggingFace Hub ID for the NER model
    llm_routing=False,      # bool — enable entropy-gated LLM routing
    llm_source="openai",    # str  — "openai" or "ollama"
    llm_model="gpt-4o",     # str  — LLM model name
    typed=True,             # bool — typed placeholders ([NAME]) vs generic ([PII])
)
result = ng.redact("Hi, I'm John Smith. Email: [e-mail]")

print(result.text)
# "Hi, I'm [NAME] [NAME]. Email: [EMAIL]"

print(result.mapping)
# {"NAME_0": "John", "NAME_1": "Smith", "EMAIL_0": "[e-mail]"}

print(result.entities)
# [{"label": "GIVENNAME", "text": "John", "confidence": 0.998, "source": "base"}, ...]

Batch:

texts = [
    "Hi, I'm John Smith. Email: [e-mail]",
    "Call me at [phone] or find me on LinkedIn.",
]

results = [ng.redact(t) for t in texts]  # model stays cached across calls

for r in results:
    print(r.text)
# "Hi, I'm [NAME] [NAME]. Email: [EMAIL]"
# "Call me at [PHONE] or find me on LinkedIn."

# Collect all mappings
all_mappings = {k: v for r in results for k, v in r.mapping.items()}
# {"NAME_0": "John", "NAME_1": "Smith", "EMAIL_0": "[e-mail]", "PHONE_0": "[phone]"}

LLM routing

Improves recall on ambiguous spans (phone numbers, IDs, dates) by routing uncertain predictions to an LLM.

# Cloud (requires OPENAI_API_KEY)
ng = Redactor(llm_routing=True, llm_source="openai", llm_model="gpt-4o")

# Local — no data leaves the machine (requires Ollama)
ng = Redactor(llm_routing=True, llm_source="ollama", llm_model="qwen2.5:7b")

CLI / interactive REPL

nerguard                                         # interactive REPL
nerguard --file report.txt                       # redact a file
nerguard --llm --backend ollama --model qwen2.5:7b  # with local LLM
nerguard --format rag                            # RAG-optimised output

REPL command	Description
/mode [human|rag|json|generic]	Switch output format
/llm	Toggle LLM routing
/backend [openai|ollama]	Switch LLM backend
/model NAME	Set LLM model
/file PATH	Redact a file
/help	Show all commands

Constructor parameters

Redactor(
    model_path=None,        # str  — local path or HuggingFace Hub ID for the NER model
    llm_routing=False,      # bool — enable entropy-gated LLM routing
    llm_source="openai",    # str  — "openai" or "ollama"
    llm_model="gpt-4o",     # str  — LLM model name
    typed=True,             # bool — typed placeholders ([NAME]) vs generic ([PII])
)

Parameter	Type	Default	Description
model_path	str	HuggingFace auto-download	Local filesystem path or HuggingFace Hub ID for the NER model. Omit to download exdsgift/NerGuard-0.3B automatically on first use.
llm_routing	bool	False	Enable entropy-gated LLM routing. When True, spans where the base model is uncertain are re-evaluated by the LLM. Improves recall on ambiguous tokens (phone numbers, dates, IDs) at the cost of extra latency.
llm_source	str	"openai"	LLM backend to use when llm_routing=True. "openai" calls the OpenAI API (requires OPENAI_API_KEY); "ollama" runs inference locally via Ollama (no data leaves the machine).
llm_model	str	"gpt-4o"	Model name passed to the selected LLM backend. Examples: "gpt-4o", "gpt-4o-mini" for OpenAI; "qwen2.5:7b", "llama3.1:8b" for Ollama. Only used when llm_routing=True.
typed	bool	True	Controls placeholder style. True → typed placeholders such as [NAME], [EMAIL], [PHONE] (preserves semantic context for downstream LLMs). False → every entity becomes [PII] regardless of type (maximum compression, no semantic signal).

RedactResult fields

ng.redact(text) returns a RedactResult dataclass with three fields:

Field	Type	Description
text	str	Redacted text with placeholders replacing PII spans.
entities	list[dict]	One dict per detected entity, with keys: label (entity type), text (original value), start/end (char offsets), confidence (0–1), source ("base" or "llm").
mapping	dict[str, str]	Maps each placeholder instance to its original value, keyed as "<LABEL>_<index>" (e.g. "NAME_0", "EMAIL_0"). Useful for auditing or selective de-redaction.

result = ng.redact("Hi, I'm John Smith. Email: [e-mail]")

result.text
# "Hi, I'm [NAME] [NAME]. Email: [EMAIL]"

result.mapping
# {"NAME_0": "John", "NAME_1": "Smith", "EMAIL_0": "[e-mail]"}

result.entities
# [
#   {"label": "GIVENNAME", "text": "John",          "start": 8,  "end": 12, "confidence": 0.998, "source": "base"},
#   {"label": "SURNAME",   "text": "Smith",         "start": 13, "end": 18, "confidence": 0.995, "source": "base"},
#   {"label": "EMAIL",     "text": "[e-mail]", "start": 27, "end": 40, "confidence": 0.991, "source": "base"},
# ]

Detected entity types

GIVENNAME · SURNAME · EMAIL · TELEPHONENUM · SOCIALNUM · CREDITCARDNUMBER · IBAN · PASSPORTNUM · IDCARDNUM · DRIVERLICENSENUM · TAXNUM · STREET · BUILDINGNUM · CITY · ZIPCODE · DATE · TIME · AGE · SEX · TITLE

Links

• Model on HuggingFace
• GitHub

License

MIT