Novel world building

Novel World-Building

Ground narrative descriptions by retrieving location and setting details from Milvus before generating environmental content.

Workflow

Before writing scenes set in specific locations:

1. Identify primary locations

   • Extract all location names from the new scene
   • Categorize as: cities, buildings, natural areas, fantasy realms
   • Note if location is first mention or recurring

2. Query Milvus for location details

   • For each location, query: "[location name] description, geography, architecture"
   • Retrieve top 8-10 chunks for comprehensive context
   • Combine results into a location profile

3. Extract environmental details From retrieved passages, document:

   • Geography: terrain, climate, distance to other locations
   • Architecture: buildings, streets, landmarks
   • Atmosphere: mood, lighting, sensory details
   • Inhabitants: who lives/works there, population characteristics
   • History: past events that shaped the location
   • Unique elements: distinctive features that make it memorable

4. Check for consistency

   • Verify climate aligns with geography
   • Confirm travel times between locations match previous mentions
   • Check that established landmarks remain in same positions
   • Validate population and infrastructure scale

5. Inject into prompt context

   • Include location profiles in system prompt before generation
   • Format as structured location data with key characteristics
   • Use specific sensory details to enrich descriptions

6. Generate environment-grounded scenes

   • Model writes scenes with retrieved location context
   • Environmental descriptions now consistent with established world
   • New location details embeddable for future retrieval

Location Profile Template

Use this format when injecting location data:

Location: [Location Name]
- Geography: [terrain, climate, cardinal position]
- Architecture: [key buildings, streets, landmarks]
- Atmosphere: [mood, lighting, sensory details]
- Inhabitants: [who's here, population type]
- History: [relevant past events]
- Travel: [distance/time to connected locations]
- Unique elements: [distinctive features]

Query Strategy

Effective Milvus queries for location retrieval:

• "[location] geography, landscape, terrain" → spatial consistency
• "[location] architecture, buildings, streets" → structural details
• "[location] weather, climate, season" → environmental conditions
• "[location] inhabitants, people, culture" → social context
• "[location] description, appears, looks" → sensory grounding
• "[location] history, past, event" → contextual depth
• "travel from [location A] to [location B]" → geographic relationships

World-Building Consistency Checks

Prevent environmental contradictions:

• Geography: locations can't move; travel times must be consistent
• Architecture: buildings once described shouldn't change appearance
• Climate: seasonal consistency within timeframe
• Population: established city sizes shouldn't drastically change
• History: past events shouldn't be retconned
• Accessibility: roads/paths established remain accessible
• Cultural markers: unique cultural elements stay consistent

Multi-Location Context

For scenes spanning multiple locations:

1. Query each location separately
2. Retrieve travel/distance information between them
3. Verify timeline feasibility (character can't be in two places simultaneously)
4. Include all location profiles in single context injection

Tips

• Query before introducing locations or during location changes
• Include surrounding areas/regions for geographic grounding
• When characters travel, retrieve both origin and destination
• Describe weather/season consistently within narrative timeframe
• After generating, embed location descriptions to enrich vector DB