Extraction Pipeline

Detailed pipeline order for both initial and event extraction, including which extractors make LLM calls and which are computed locally.

Initial Extraction Pipeline

Runs on the first message to create the initial snapshot.

#	Extractor	Category	LLM Call?	Dependencies
1	Time	`time`	Yes (unless card default)	None
2	Location	`location`	Yes (unless card default)	None
3	Forecast	`climate`	No (computed)	Location + Time
4	Characters Present	`characters`	Yes	None
5	Character Profiles	`characters`	Yes (per character)	Characters Present
6	Character Outfits	`characters`	Yes (per character)	Characters Present
7	Relationships	`relationships`	Yes	Characters Present
8	Props	`props`	Yes	Location + Character Outfits
9	Topic & Tone	`scene`	Yes	None
10	Tension	`scene`	Yes	Topic & Tone

Total LLM calls: 6 + (2 × number of characters) minimum

For example, with 2 characters: 6 + 4 = 10 LLM calls for initial extraction.

Card defaults can reduce this:

Time from card → skips step 1 (-1 call)
Location from card → skips step 2 (-1 call)

Event Extraction Pipeline

Runs on every subsequent message. Organized in phases.

Phase 1: Core State

Extractor	Category	LLM Call?	Notes
Time Change	`time`	Yes
Location Change	`location`	Yes
Forecast	`climate`	No (computed)	Triggers on area change or time exceeding range
Topic/Tone Change	`scene`	Yes
Tension Change	`scene`	Yes

Phase 2: Character Presence

Extractor	Category	LLM Call?	Notes
Presence Change	`characters`	Yes	Detects appeared/departed

Phase 3: Per-Character State

Runs once per present character:

Extractor	Category	LLM Call?	Notes
Position/Activity Change	`characters`	Yes	Combined in one call
Mood Change	`characters`	Yes
Outfit Change	`characters`	Yes

Phase 4: Props

Extractor	Category	LLM Call?	Notes
Props Change	`props`	Yes	After outfits (clothing ↔ props)
Props Confirmation	`props`	Yes	Confirms props still present

Phase 5: Relationship Subjects (Global)

Extractor	Category	LLM Call?	Notes
Subject Detection	`relationships`	Yes	Identifies interaction types

Phase 6: Per-Pair Relationships

Runs once per character pair:

Extractor	Category	LLM Call?	Notes
Relationship Change	`relationships`	Yes	Updates feelings/wants/secrets/status

Phase 7: Narrative

Extractor	Category	LLM Call?	Notes
Narrative Description	`narrative`	Yes	Brief summary of what happened
Milestone Description	`narrative`	Yes (conditional)	Only for first-occurrence milestone subjects

Phase 8: Chapters

Extractor	Category	LLM Call?	Notes
Chapter Ended	`narrative`	Yes	Detects chapter boundaries
Chapter Description	`narrative`	Yes (conditional)	Only when chapter ends

Total LLM Calls per Message

With all modules enabled, N characters, and P pairs (P = N×(N-1)/2):

Core:           4  (time + location + topic/tone + tension)
Presence:       1
Per-character:  3N (position/activity + mood + outfit per character)
Props:          2  (change + confirmation)
Subjects:       1
Per-pair:       P
Narrative:      1-2 (description + optional milestone)
Chapters:       1-2 (detection + optional description)
────────────────────
Total:          10 + 3N + P  (minimum, all modules on)

Examples:

2 characters, 1 pair: 10 + 6 + 1 = 17 calls
3 characters, 3 pairs: 10 + 9 + 3 = 22 calls
4 characters, 6 pairs: 10 + 12 + 6 = 28 calls

Reducing Call Count

Disable modules you don’t need:

Disable	Calls Saved
`props`	2 per message
`scene` (topic/tone + tension)	2 per message
`narrative` (events + chapters)	2-4 per message
`relationships`	1 + P per message
`characters`	1 + 3N per message

shouldRun Logic

Not every extractor runs every time. Each has a shouldRun check:

Track disabled → extractor is skipped entirely
Run strategy → some extractors only run on certain conditions (e.g., forecast only when area changes)
No relevant changes → some extractors check if upstream state actually changed

Computed vs Extracted

Some state is computed locally rather than extracted by the LLM:

State	Source
Climate (temperature, conditions, etc.)	Computed from forecast + time + location
Forecast	Computed from climate normals + seeded RNG
Daylight phase	Computed from latitude + time
Indoor temperature	Computed from outdoor temp + building type
Narrative event witnesses	Derived from projection state at each message
Chapter boundaries (partial)	Detected from location changes + time jumps

Event Types Track Dependencies