piighost 0.4.2

PII anonymization middleware for AI agent conversations using LangChain integration.

PIIGhost

piighost is a Python library that detects PII (personally identifiable information), extracts them, applies corrections, and automatically anonymizes and deanonymizes sensitive entities (names, locations, etc.). With modules for bidirectional anonymization in AI agent conversations, it integrates via a LangChain middleware without modifying your existing agent code.

Features

• Detection: Detect PII with NER models, algorithms, and build your custom configuration with our detector composition component
• Span resolution: Resolve overlapping or nested detected spans to guarantee clean, non-redundant entities, especially when using multiple detectors
• Entity linking: Link different detections together, enabling typo tolerance and catching mentions that an NER model might miss
• Entity resolution: Resolve linked entity conflicts (e.g., one detector links A and B, another links B and C) to guarantee coherent final entities
• Anonymization: Anonymize detected entities with customizable placeholders (e.g., <<PERSON_1>>, <<LOCATION_1>>) to protect privacy while preserving text structure. A cache system remembers the applied anonymization and can reverse it for deanonymization
• Placeholder Factory: Create custom placeholders for anonymization, with flexible naming strategies (counters, UUID, etc.) to fit your specific needs
• Middleware: Easily integrate piighost into your LangChain agents for transparent anonymization before and after model calls, without modifying your existing agent code

Installation

Basic installation

This project uses uv for dependency management.

uv add piighost
uv pip install piighost

Development installation

Clone the repository and install with dev dependencies:

git clone https://github.com/Athroniaeth/piighost.git
cd piighost
uv sync

Makefile helpers

Run the full lint suite with the provided Makefile:

make lint

This runs Ruff (format + lint) and PyReFly (type-check) through uv run.

Quick start

Standalone pipeline

import asyncio

from piighost.anonymizer import Anonymizer
from piighost.detector.gliner2 import Gliner2Detector
from piighost.linker.entity import ExactEntityLinker
from piighost.resolver import MergeEntityConflictResolver, ConfidenceSpanConflictResolver
from piighost.pipeline import AnonymizationPipeline
from piighost.placeholder import CounterPlaceholderFactory

from gliner2 import GLiNER2

entity_linker = ExactEntityLinker()
entity_resolver = MergeEntityConflictResolver()
span_resolver = ConfidenceSpanConflictResolver()

ph_factory = CounterPlaceholderFactory()
anonymizer = Anonymizer(ph_factory=ph_factory)

model = GLiNER2.from_pretrained("urchade/gliner_multi-v2.1")
detector = Gliner2Detector(
    model=model,
    threshold=0.5,
    labels=["PERSON", "LOCATION"],
)

pipeline = AnonymizationPipeline(
    detector=detector,
    span_resolver=span_resolver,
    entity_linker=entity_linker,
    entity_resolver=entity_resolver,
    anonymizer=anonymizer,
)


async def main():
    text = "Patrick lives in Paris. Patrick loves Paris."
    anonymized, entities = await pipeline.anonymize(text)
    print(anonymized)
    # <<PERSON_1>> lives in <<LOCATION_1>>. <<PERSON_1>> loves <<LOCATION_1>>.

    original, _ = await pipeline.deanonymize(anonymized)
    print(original)
    # Patrick lives in Paris. Patrick loves Paris.


asyncio.run(main())

With LangChain middleware

from langchain.agents import create_agent
from langchain_core.tools import tool

from piighost.anonymizer import Anonymizer
from piighost.detector.gliner2 import Gliner2Detector
from piighost.linker.entity import ExactEntityLinker
from piighost.resolver import MergeEntityConflictResolver, ConfidenceSpanConflictResolver
from piighost.pipeline import ThreadAnonymizationPipeline
from piighost.placeholder import CounterPlaceholderFactory
from piighost.middleware import PIIAnonymizationMiddleware

from gliner2 import GLiNER2


@tool
def send_email(to: str, subject: str, body: str) -> str:
    """Send an email to a given address."""
    return f"Email successfully sent to {to}."


entity_linker = ExactEntityLinker()
entity_resolver = MergeEntityConflictResolver()
span_resolver = ConfidenceSpanConflictResolver()

ph_factory = CounterPlaceholderFactory()
anonymizer = Anonymizer(ph_factory=ph_factory)

model = GLiNER2.from_pretrained("urchade/gliner_multi-v2.1")
detector = Gliner2Detector(
    model=model,
    threshold=0.5,
    labels=["PERSON", "LOCATION"],
)
pipeline = ThreadAnonymizationPipeline(
    detector=detector,
    span_resolver=span_resolver,
    entity_linker=entity_linker,
    entity_resolver=entity_resolver,
    anonymizer=anonymizer,
)
middleware = PIIAnonymizationMiddleware(pipeline=pipeline)

graph = create_agent(
    model="openai:gpt-5.4",
    system_prompt="You are a helpful assistant.",
    tools=[send_email],
    middleware=[middleware],
)

The middleware intercepts every agent turn the LLM only sees anonymized text, tools receive real values, and user-facing messages are deanonymized automatically.

How it works

Anonymization pipeline

---
title: "piighost AnonymizationPipeline.anonymize() flow"
---
flowchart LR
    classDef stage fill:#90CAF9,stroke:#1565C0,color:#000
    classDef protocol fill:#FFF9C4,stroke:#F9A825,color:#000
    classDef data fill:#A5D6A7,stroke:#2E7D32,color:#000

    INPUT(["`**Input text**
    _'Patrick lives in Paris.
    Patrick loves Paris.'_`"]):::data

    DETECT["`**1. Detect**
    _AnyDetector_`"]:::stage
    RESOLVE_SPANS["`**2. Resolve Spans**
    _AnySpanConflictResolver_`"]:::stage
    LINK["`**3. Link Entities**
    _AnyEntityLinker_`"]:::stage
    RESOLVE_ENTITIES["`**4. Resolve Entities**
    _AnyEntityConflictResolver_`"]:::stage
    ANONYMIZE["`**5. Anonymize**
    _AnyAnonymizer_`"]:::stage

    OUTPUT(["`**Output**
    _'<<PERSON_1>> lives in <<LOCATION_1>>.
    <<PERSON_1>> loves <<LOCATION_1>>.'_`"]):::data

    INPUT --> DETECT
    DETECT -- "list[Detection]" --> RESOLVE_SPANS
    RESOLVE_SPANS -- "deduplicated detections" --> LINK
    LINK -- "list[Entity]" --> RESOLVE_ENTITIES
    RESOLVE_ENTITIES -- "merged entities" --> ANONYMIZE
    ANONYMIZE --> OUTPUT

    P_DETECT["`GlinerDetector
    _(GLiNER2 NER)_`"]:::protocol
    P_RESOLVE_SPANS["`ConfidenceSpanConflictResolver
    _(highest confidence wins)_`"]:::protocol
    P_LINK["`ExactEntityLinker
    _(word-boundary regex)_`"]:::protocol
    P_RESOLVE_ENTITIES["`MergeEntityConflictResolver
    _(union-find merge)_`"]:::protocol
    P_ANONYMIZE["`Anonymizer + CounterPlaceholderFactory
    _(<<LABEL_N>> tags)_`"]:::protocol

    P_DETECT -. "implements" .-> DETECT
    P_RESOLVE_SPANS -. "implements" .-> RESOLVE_SPANS
    P_LINK -. "implements" .-> LINK
    P_RESOLVE_ENTITIES -. "implements" .-> RESOLVE_ENTITIES
    P_ANONYMIZE -. "implements" .-> ANONYMIZE

Each stage uses a protocol (structural subtyping) swap GlinerDetector for spaCy, a remote API, or an ExactMatchDetector for tests. Same for every other stage.

Middleware integration

---
title: "piighost PIIAnonymizationMiddleware in an agent loop"
---
sequenceDiagram
    participant U as User
    participant M as Middleware
    participant L as LLM
    participant T as Tool

    U->>M: "Send an email to Patrick in Paris"
    M->>M: abefore_model()<br/>NER detect + anonymize
    M->>L: "Send an email to <<PERSON_1>> in <<LOCATION_1>>"
    L->>M: tool_call(send_email, to=<<PERSON_1>>)
    M->>M: awrap_tool_call()<br/>deanonymize args
    M->>T: send_email(to="Patrick")
    T->>M: "Email sent to Patrick"
    M->>M: awrap_tool_call()<br/>reanonymize result
    M->>L: "Email sent to <<PERSON_1>>"
    L->>M: "Done! Email sent to <<PERSON_1>>."
    M->>M: aafter_model()<br/>deanonymize for user
    M->>U: "Done! Email sent to Patrick."

Development

uv sync                      # Install dependencies
make lint                    # Format (ruff), lint (ruff), type-check (pyrefly)
uv run pytest                # Run all tests
uv run pytest tests/ -k "test_name"  # Run a single test

Contributing

• Commits: Conventional Commits via Commitizen (feat:, fix:, refactor:, etc.)
• Type checking: PyReFly (not mypy)
• Formatting/linting: Ruff
• Package manager: uv (not pip)
• Python: 3.12+

Additional notes

• The GLiNER2 model is downloaded from HuggingFace on first use (~500 MB)
• All data models are frozen dataclasses safe to share across threads
• Tests use ExactMatchDetector to avoid loading the real GLiNER2 model in CI