Glossary¶

Key terms used throughout the piighost documentation. Keep this page open if you are new to the library or to NER.

PII

Personally Identifiable Information. Any piece of data that can identify a person: name, address, phone number, email, location, organization, account number. piighost detects and anonymizes PII so that third-party LLMs never see the raw values.

LLM

Large Language Model. A neural network trained on large text corpora to generate or reason about text (GPT, Claude, Gemini, Mistral, etc.). In this project, an LLM is the downstream consumer that should only receive anonymized input.

NER

Named Entity Recognition. A machine-learning task that finds named entities (persons, locations, organizations, dates, etc.) in text. piighost ships detectors for several backends (SpacyDetector, Gliner2Detector, TransformersDetector), all interchangeable behind the AnyDetector protocol.

Detector

A component that finds PII in text. Detectors implement the AnyDetector protocol. Built-in implementations: ExactMatchDetector (fixed dictionary), RegexDetector (patterns), GlinerDetector (NER), CompositeDetector (chain of detectors).

Span

An interval of character positions inside a text: (start_pos, end_pos). Each detection carries a Span to pinpoint where the PII appears. Overlapping spans from multiple detectors are resolved by the span resolver.

Detection

The output of a detector: a (text, label, span, confidence) tuple. For example, detecting Patrick as PERSON at position (0, 7) with confidence 0.95 produces one Detection.

Entity

A logical PII that may appear multiple times in the text. Produced by the entity linker by grouping related detections (typo variants, case variants, partial mentions). Different from a Detection, which is one spotted occurrence.

Entity linking

The step that groups detections referring to the same real-world PII. For example, linking Patrick at position (0, 7) and patrick at position (34, 41) into one Entity, so both occurrences share the same placeholder.

Placeholder

The token that replaces a PII in the anonymized text. Defaults look like <<PERSON:1>>, <<LOCATION:1>>. The naming strategy is controlled by a PlaceholderFactory (counter-based, UUID, hashed, redacted, masked).

Pipeline

The 5-stage orchestration that transforms raw text into anonymized text: Detect, Resolve Spans, Link Entities, Resolve Entities, Anonymize. Implemented by AnonymizationPipeline (stateless) and ThreadAnonymizationPipeline (conversation-scoped).

Resolver

A component that arbitrates conflicts. Two kinds: SpanConflictResolver (overlapping detections) and EntityConflictResolver (linked entity groups that share a mention).

Middleware

A LangChain extension point that runs before and after every LLM call and every tool call. PIIAnonymizationMiddleware hooks into it to intercept and transform messages, so anonymization applies without changing agent code.

Thread

A conversation scope identified by a thread_id. Memory and cache are isolated per thread so two parallel conversations do not share PII state.