Extraction Lifecycle

BlazeTracker has two distinct extraction paths: initial extraction for the first message in a chat, and event extraction for all subsequent messages. Understanding when each runs and what they produce is key to understanding how state flows through the system.

When Extraction Triggers

Extraction can be triggered in several ways:

Auto-extract is controlled by the Auto Extract setting. When enabled, extraction runs automatically whenever a new assistant message is received.

Initial Extraction

Initial extraction runs on the first assistant message and produces a Snapshot — a complete state capture that serves as the starting point for all future projections.

The initial pipeline runs extractors in a specific order because some depend on previous results:

Time → Location → Forecast → Characters Present → Character Profiles
→ Character Outfits → Relationships → Props → Topic & Tone → Tension

Key details:

• Each extractor receives the partial snapshot built by previous extractors
• Forecast needs both location and time to generate weather
• Props runs after outfits to filter worn clothing from nearby objects
• Character Profiles and Outfits run per-character (one LLM call each)
• Relationships runs once with all present characters
• If a card extension provides absolute values (e.g., starting time, starting location), the corresponding LLM call is skipped entirely

Event Extraction

Event extraction runs on every subsequent message. Instead of building a complete snapshot, it detects what changed and produces discrete events.

The event pipeline runs in phases:

Phase 1: Core State

Time Change → Location Change → Forecast → Topic/Tone Change → Tension Change

These are global extractors that run once per message.

Phase 2: Character Presence

Presence Change (appeared/departed)

Detects characters entering or leaving the scene.

Phase 3: Per-Character State

For each present character:
  Position/Activity Change → Mood Change → Outfit Change

These run per character — if 3 characters are present, that’s 3 rounds of LLM calls for each extractor type.

Phase 4: Props

Props Change → Props Confirmation

Runs after outfit changes so removed clothing can become scene props and picked-up props can be removed.

Phase 5: Relationship Subjects

Subject Detection (global)

Identifies what interaction types occurred (conversation, flirt, argument, etc.).

Phase 6: Per-Pair Relationships

For each character pair:
  Relationship Change

Updates feelings, secrets, wants, and status for each pair.

Phase 7: Narrative

Narrative Description → Milestone Description

Generates a brief summary of what happened and, for first-occurrence milestone subjects, a milestone description.

Phase 8: Chapters

Chapter Ended Detection → Chapter Description

Checks if a chapter boundary occurred (location change or time jump) and generates a title/summary if so.

Abort Handling

Every phase checks for an abort signal before running. If the user stops generation or clicks the 🔥 button during extraction, the pipeline aborts cleanly:

• Events already produced in earlier phases are not saved — it’s all-or-nothing per turn
• The abort is detected via an AbortController signal
• The UI shows the extraction as incomplete

shouldRun Logic

Not every extractor runs on every message. Each extractor has a shouldRun check:

• Track settings — If a module is disabled (e.g., track.climate = false), its extractors are skipped
• Run strategy — Some extractors only run every N turns or when specific conditions are met
• Dependencies — If a required extractor produced no results, dependent extractors may skip

This means the actual number of LLM calls per message varies depending on your configuration and what changed in the narrative.