Entity Detection & Mentions

As you write in the content editor, MythTapestry can detect references to entities in your text and suggest linking them to your knowledge graph. This feature is entirely optional — you can always link entities manually using the @ mention system instead.

How it works

1. You write content that mentions an entity (e.g. “Aldric rode to the Crimson Tower”)
2. The system detects “Aldric” and “Crimson Tower” as potential entity references
3. Suggestions appear as highlighted text with a subtle indicator
4. You can confirm or dismiss each suggestion

Detection methods

Two methods are available (configured in World Settings):

GLiNER (local, no AI)

GLiNER is a named-entity recognition model — it is not a large language model (LLM). It runs entirely on the MythTapestry server using pattern-matching and lightweight neural inference, with no external API calls.

• Fast — typically processes text in under a second
• No quota usage and no data sent to external services
• Good for common entity types (names, places, organisations)
• Uses fuzzy matching (via rapidfuzz) against your world’s entity names, aliases, and subtypes
• May miss domain-specific or context-dependent entities

Data privacy — GLiNER mode

What’s processed: Your content text and your world’s entity names, types, and aliases.

Where it runs: Entirely on the MythTapestry server. No data is sent to any external AI provider. The GLiNER model (GLiNER-medium-v2.1) runs locally alongside the application.

Planned: A future client-side GLiNER mode is planned that would run the model directly in your browser, so that text never leaves your device at all.

AI-powered

Uses your configured AI model via OpenRouter for more nuanced detection. This is better for unusual references and context-dependent names (e.g. recognising “The Crown” as a faction rather than a literal crown).

• More accurate for unusual or context-dependent references
• Uses your feature quota
• Better at understanding when common words are entity names

Data privacy — AI mode

What’s sent: The plain text of the content being analysed, plus the names, types, and aliases of entities in your world (so the model knows what to look for).

What’s NOT sent: Your user identity, session IDs, world ID, or any metadata beyond what’s needed for detection. No images, files, or relationship data are included.

Where it goes: To the AI model provider via OpenRouter, as configured in your model preferences.

Detection modes

Mode	Trigger	Description
Manual	Click the “Detect Entities” button	You control when detection runs
Auto	On content save	Detection runs automatically when you save

The mode is configured per world in World Settings. You can disable detection entirely if you prefer to manage mentions manually.

Managing mentions

Single mention

When one entity is detected in a text selection, you see an inline confirm/dismiss action right at the mention location.

Multiple mentions

When several entities are detected, they appear in a collapsible panel with:

• Bulk confirm — accept all suggestions at once
• Bulk dismiss — reject all suggestions
• Individual confirm/dismiss per mention

Tip

You can also manually mention an entity by typing @ in the editor and searching for it. Manual mentions are always confirmed immediately — no detection system needed.

Mention states

State	Appearance	Behaviour
Confirmed	Solid highlight, clickable link	Persists across saves, visible in knowledge graph
Suggested	Subtle highlight, action buttons	Survives saves but doesn’t create graph links
Dismissed	Removed	Won’t be suggested again for this text

Note

The mention sync system preserves unconfirmed suggestions when you save. Only confirmed mentions create actual entity links in the knowledge graph — suggestions are stored as pending records that don’t affect your graph until confirmed.

New entity suggestions

When enabled (suggest_new_entities in World Settings), the detection system can flag text that looks like it refers to an entity that doesn’t exist yet in your world. These appear with a special indicator so you can decide whether to create the entity.

Confidence threshold

The confidence threshold (default 0.7) controls how certain the detection must be before suggesting a mention. Lower values produce more suggestions but may include false positives. Higher values are more conservative.