LLM World Engine Knowledge Base

User-veritasr

6 min read

veritasr

Overview

Primary Contributor: Most active participant in the channel Primary Project: ReallmCraft (LLM World Engine) Key Innovation: Natural Description Language (NDL) Technical Focus: Backend architecture, RAG systems, constraint programming Active Period: January 2024 - July 2025 (throughout entire transcript)

Background & Approach

veritasr brought significant software engineering experience to the project, often referencing day-job practices:

  • Forced to provide 3 separate solutions during design phase at work
  • Strong emphasis on modularity and separation of concerns
  • Preference for programmatic control over LLM decision-making

Philosophy

“Turns out that when you take away decision making from the LLM it behaves much better.”

Major Contributions

1. ReallmCraft Engine

Architecture: Flask-based backend with React frontend Key Features:

  • Modular world generation
  • Template-based entity system
  • RAG retrieval for context management
  • Custom scripting language for game logic
  • Mod framework support

Development Timeline:

  • Jan 2024: Initial architecture discussions
  • Feb 2024: Core engine development, world management
  • Feb-Jun 2024: Feature expansion, UI development
  • Jun 2024+: NDL development, constraint programming exploration
  • Jul 2025: Advanced prompting workflows

Code Scale: ~7,500+ lines reported (July 2024)

2. Natural Description Language (NDL)

Purpose: Convert programmatic game events into natural narrative

Core Concept:

Backend Events/State → NDL Markup → LLM Processing → Natural Text

Philosophy:

“Essentially what I’m doing is that I’m simulating events and state changes on the backend. Those events get passed into a process that converts the events to a markup language, effectively building a prompt, which is then passed to the LLM as a decorator.”

Key Insight: LLM functions as translator, not decision-maker

Results Demonstrated (July 2025):

  • Consistent subtext in narrative
  • Controlled emotional beats
  • Reliable dialogue with implicit intent
  • Works across multiple models (Gemma, Llama3, Mistral, Stheno)

3. RAG & Memory Systems

Approach: Hybrid retrieval combining multiple methods

Implementation:

  • Exact name matching for direct references
  • Tag-based boolean search
  • Semantic similarity for descriptions
  • Ranking and fusion (RRF-style)
  • Context window management with token estimation

Key Innovation: Treating semantic search as paragraph comparison, not keyword matching

“Semantic similarity isn’t meant to locate stuff like keywords or key phrases, it’s meant to find similar paragraphs of text.”

Solution: Store examples of experiences rather than

descriptions

4. Constraint Programming Integration

Discovery (July 2024): Realized probability/statistics might be wrong approach

New Paradigm:

  • Scoring expands, probability contracts
  • Constraint-based reasoning (deduction)
  • Rule engines (induction)
  • Utility systems (weighing options)

Four Component Types:

  1. Generators - Create and cache options/choices
  2. Assemblers - Compose elements from choices, build objects/state
  3. Evaluators - Parse conditions, score, make choices
  4. Analyzers - Abstract data, derive metrics from existing data

5. Advanced Prompting Techniques

Late-Stage Development (July 2025):

Dialogue System:

Character beliefs → Never state outright
Rules about what can/cannot be said
Information that comes to mind
What they wish to convey vs. what they do convey

Results: Achieved natural subtext and implicit character motivation without explicit dialogue scripting

Narrative Generation:

  • Event-based narration
  • Subtext through omission and implication
  • Controlled emotional beats
  • Works with smaller models (7B-9B)

Technical Expertise

Backend Development

  • Python (Flask, FastAPI patterns)
  • PostgreSQL/SQLite
  • REST API design
  • Background task processing

Frontend Development

  • React/Next.js
  • Considered Tauri for desktop deployment
  • UI/UX for complex systems

AI/ML

  • RAG architectures
  • Vector databases (ChromaDB, Qdrant)
  • Prompt engineering
  • Multi-model support (Ooba, Kobold, OpenAI, OpenRouter, Mancer)

System Design

  • Modular architecture
  • Template systems
  • Constraint programming
  • Rule engines
  • Event-driven systems

Development Philosophy

Core Principles

  1. Programmatic Control Over LLM Decision-Making

    • State changes must be deterministic
    • LLM for narration only
    • Separate logic from presentation

  2. Data-Driven Design

    • Everything configurable via templates
    • JSON-based entity definitions
    • Modding support from ground up

  3. Just-In-Time Generation

    • Don’t create everything upfront
    • Generate as needed
    • Fix values once created

  4. Modularity & Separation of Concerns

    • Backend vs. Frontend separation
    • Logic vs. Narrative layers
    • Core vs. Mod systems

  5. Iterative Development

    • “I’m like on iteration 12 for this thing”
    • Willing to throw out and rebuild
    • Test-driven refinement

Interaction Style

Collaborative

  • Shares detailed technical implementations
  • Provides code examples freely
  • Debugs others’ issues
  • Answers questions comprehensively

Experimental

  • Constantly trying new approaches
  • Documents findings
  • Pivots when necessary
  • “Think tank” mentality

Pragmatic

  • “Better result would be something like: ‘I’m not sure, but this is the direction I was thinking…‘”
  • Focuses on what works over theoretical purity
  • Balances ideal vs. practical

Self-Aware

  • Acknowledges when taking over thread
  • Apologizes for walls of text
  • Transparent about iteration count
  • Admits limitations

Notable Quotes

On LLM Limitations

“LLMs are terrible at state management - Keep state programmatic”

On Architecture

“Essentially half simulation half story”

On Development

“A lot of these ideas make it to partial implementations before throwing them out”

On Sharing Work

“I don’t typically share projects, since that means people want me to support them. I’ve been burned by that in the past.”

On LLM Role

“Sort of like it just narrating what happens in the world rather than the LLM trying to continue the story”

On Innovation

“It sort of feels like the industry is catching up to what we were doing in here last year”

Timeline of Major Developments

Early 2024 (Jan-Mar)

  • Joined thread, began ReallmCraft development
  • Established core architecture principles
  • Built world generation system
  • Implemented RAG retrieval

Mid 2024 (Apr-Jun)

  • Frontend development (React)
  • Micromamba integration for cross-platform support
  • Workflow systems
  • Random table generation
  • Extensive playtesting (Zweihander-inspired world)

Late 2024 (Jul-Dec)

  • Character generation systems
  • Behavior systems
  • NDL initial development
  • Constraint programming exploration
  • Modding framework

2025 (Jan-Jul)

  • NDL refinement
  • Advanced prompting workflows
  • Selection-based UI (reducing typing fatigue)
  • Staged/contextual menus (FSM-based)
  • Subtext and implicit dialogue systems

Project Status

ReallmCraft:

  • Not publicly released on GitHub (as of analysis)
  • Went through ~12 iterations
  • Current iteration: Modernized expert system with LLM as decorator
  • “Very WIP… like 2 weeks in or something like that” (re: latest iteration)

NDL:

  • Functional and tested across multiple models
  • Achieves reliable narrative generation
  • Successfully handles subtext and character motivation
  • Works with small models (7B-9B parameters)

Lessons Shared

  1. On Context Management:

    • Token estimation with padding (1/50 ratio)
    • Priority-based context inclusion
    • Exact matches first, semantic search second

  2. On Content Generation:

    • LLM-generated templates work well
    • Cache and refine
    • Focus tooling on data creation
    • Consider crowdsourcing content

  3. On Retrieval:

    • Three approaches: RAG, boolean (keyword), knowledge graphs
    • Hybrid works best
    • Store experiences, not descriptions
    • Generate hypothetical questions for better matching

  4. On Architecture:

    • Backend work easier than frontend (for them)
    • Security concerns only matter for multi-user
    • Local-first design simplifies considerably
    • Plan for modding from start

Technical Challenges Overcome

  1. Cross-Platform Support: Micromamba/conda integration for Windows
  2. Context Window Management: Dynamic token estimation and pruning
  3. Retrieval Quality: Hybrid boolean + semantic + exact matching
  4. UI Complexity: Selection-based menus to reduce typing fatigue
  5. LLM Consistency: NDL system for controlled generation
  6. State Synchronization: Event-driven architecture with deterministic outcomes

Influence on Community

  • Dominated technical discussions (sheer volume of contributions)
  • Set standards for architectural thinking
  • Provided working code examples
  • Debugged others’ implementations
  • Pushed boundaries of what’s considered possible with LLMs

Analysis Note

veritasr represents the most prolific technical contributor in the transcript, with contributions spanning the entire 18-month period. Their work demonstrates deep systems thinking, willingness to iterate radically, and a pragmatic approach to leveraging LLMs as tools rather than solutions.

Graph View

Backlinks