Skip to content

🛡️ Utility Core

The Utility Core is the shared infrastructure layer powering Vespera's everyday capabilities. It manages personality consistency, memory budgeting, language processing, and automated moderation.

VesperaPersonality — Singleton

All 15+ cogs import their prompts, embed colors, and error messages from a single VesperaPersonality static class. This prevents Personality Drift — where the bot sounds like a corporate assistant in one command and a pirate in another.

class VesperaPersonality:
    NAME = "Vespera"
    TITLE = "The Silent Architect"

    class Colors:
        PRIMARY = 0x2C2F33   # Dark grey (base)
        SUCCESS = 0x57F287   # Muted green
        ERROR   = 0xED4245   # Red
        MYSTIC  = 0x9B59B6   # Purple (D&D)
        CLOUD   = 0x38BDF8   # Sky blue (Cloud)

Changing Vespera's tone requires editing one file — the change propagates instantly across all modules.

🌍 Translator

Status: Lazy Glossary Optimized · Personality Integrated

AI-powered translation with deep awareness of D&D and Cloud Engineering terminology.

Translation Styles

Style Description
Formal Professional, business-appropriate tone
Informal Casual conversation style
Slang Internet slang and idioms
Lyrical Poetic or dramatic — great for D&D narration

Smart Glossaries

  • D&D Glossary: Fireball, Saving Throw, Advantage, Concentration...
  • Cloud Glossary: Kubernetes, Docker, Serverless, VPC, IAM...

Triggers

  • Slash command: /translate <text> <language> <style>
  • Reaction trigger: Add a flag emoji (🇺🇸 🇯🇵 🇩🇪) to any message to auto-translate it

Technical Highlight — Lazy Glossary Injection

95% Prompt Size Reduction

The glossary is NOT injected by default. The system scans the input text for known technical keywords first. Only terms present in the input are added to the prompt.

Without Lazy Injection With Lazy Injection
~2,000 token glossary sent every time ~100 tokens of relevant terms only
RAM: ~60MB RAM: ~25MB

📝 TL;DR

Status: JSON Optimized · String Interned

Instant summarization of Discord chat logs or uploaded text files, optimized for low-memory environments.

Modes

Mode Output
Summarize General overview paragraph
Action Items Extracted to-do list and assignments
Key Points Bulleted list of critical information

Input Support

  • Last N messages from Discord chat history
  • Uploaded .txt, .md, or .log files

Technical Highlight — String Interning

Memory Optimization Detail

When summarizing a long chat log, usernames and common headers repeat thousands of times. The TL;DR module uses sys.intern() on repeated strings, ensuring that if "User_Kazeyi" appears 1,000 times, only one string object exists in memory — not 1,000 copies.

Large files are also processed in stream-safe chunks to avoid loading the entire file into RAM at once.

🛡️ Moderator

Status: Vespera Integrated

Automated threat analysis using AI sentiment and context scoring — far beyond simple keyword REGEX.

How it works

  1. The message and surrounding context are passed to the AI with a structured [Reasoning] block requirement
  2. The model must reason through: "User said X, context is Y, therefore hostility is..." before committing to a score
  3. A Hostility Score (0–10) is returned and acted upon

Technical Highlight — Chain-of-Thought for Moderation

Why CoT matters here

If a model is allowed to start its response with "This message is safe", it is statistically biased to justify that conclusion — even when wrong (false negatives).

By forcing a [Reasoning] block first, the model must step through its logic before committing to a verdict, significantly reducing both false positives and false negatives in moderation.

Optimizations

  • LRU Cache: Analysis results for frequently repeated phrases are cached (max 128 entries), avoiding redundant API calls for known-safe content
  • Personality Integration: All violation embeds use VP.Colors.ERROR and Vespera's firm but professional tone — no robotic automated-sounding warnings