How Our Models Are Developed

Overview

Ophraxx AI is currently in active beta development. The Ophraxx Web model network is our platform-wide lineup of AI — built for web users and Discord communities, designed around speed, safety, and honest, accurate responses. This page describes the model lineup, how queries are routed between models, the safety systems in place, and the rate limits that govern usage.

All models run on the Ophraxx Web Core, our proprietary AI infrastructure optimized for extremely low inference latency — an important requirement for a chat environment where users expect near-instant replies. Different model IDs represent different tuning profiles and capability levels across our subscription tiers.

The Ophraxx Web Model Network

The Ophraxx Web network is our core model lineup, powered by the Ophraxx Web Core. Each tier is designed for a different use case and capability level. All tiers share the same safety stack, system prompt, and personality layer — the tuning profile, token ceiling, and feature set differ by tier.

Status legend: In Testing (active development), Not Started (planned, not yet built). All models run on the Ophraxx Web Core.

Automatic Query Routing

Users never need to manually select a model. Ophraxx AI includes a prompt classifier that automatically routes every query to the appropriate model tier based on the subscription plan and the complexity of the request.

The classifier scores each message against two pattern sets. Complex indicators include keywords like "analyze", "explain", "algorithm", "write a", "compare", "step-by-step", "essay", "code", and academic subject names such as mathematics, physics, philosophy, and economics. Simple indicators include greetings, identity questions, and filler phrases.

Message length also feeds into the score. Any message over 180 characters is treated as complex automatically. Messages between 80 and 180 characters with qualifying keywords or multiple questions score higher toward the complex threshold. A complexity score of 2 or higher routes to the higher-tier model available on the user's plan (OW-B12 or above when live); everything else routes to the base tier model (OW-F1).

If the target model is not yet live, the classifier falls back gracefully to the base model — ensuring users always receive a response. Additionally, short acknowledgment messages ("ok", "thanks", "lol") are caught before the classifier runs and receive lightweight static replies without any model call.

Personality Modes

Server administrators can configure the tone and communication style of Ophraxx AI through five personality modes. Personality affects how the model communicates — it does not change the safety rules, which are enforced independently and cannot be overridden by any personality setting.

• Default

  Balanced, helpful, and intelligent. Naturally adapts to the tone of the conversation while remaining professional and clear. The starting mode for all servers.

• Professional

  Formal and business-focused. Uses precise, structured language. Avoids slang, casual expressions, and humor. Leads with the most important information in every response.

• Friendly

  Warm, approachable, and conversational. Uses a relaxed tone and is comfortable with light humor. Encourages and supports users like a knowledgeable friend.

• Tutor

  Patient and educational. Breaks down complex topics into clear steps, uses analogies and real-world examples, and actively invites follow-up questions. Designed for learning-focused communities.

• Concise

  Extremely direct. One to three sentences unless absolute complexity demands more. Bullet points over paragraphs. No preamble, no filler, no over-explanation. Best for fast-paced communities.

Safety Architecture

Every message passes through a multi-layer safety pipeline before reaching the model, and every response passes through a validation pipeline before being sent. The five layers are:

1. Pattern-based pre-screening

   Fast automated matching against nine content categories — sexual content, violence instructions, hate speech, self-harm expressions, illegal activity, and others. Catches obvious violations without spending a model call. Violations are logged in the user's moderation history.

2. LLM safeguard (input)

   A dedicated hosted safeguard model screens every input against the same nine threat categories (S1–S9), including prompt injection (S8) and inappropriate persona assignment (S9). Returns a precise category label rather than a vague flagged or unflagged signal.

3. Automated fact-checking

   After the main model generates a response, a secondary model checks it for factual accuracy against verifiable claims — dates, statistics, scientific facts, and technical details. Responses with clear factual errors are blocked and not sent. Responses with low-confidence claims are logged as warnings.

4. LLM safeguard (output)

   The same safeguard model screens the AI-generated response before delivery, catching any unsafe content the main model may have produced despite the system prompt constraints.

5. Output sanitization and PII redaction

   A final pass strips PII — phone numbers, email addresses, and similar identifiers — from responses before they are sent. Output is also checked against hard-block patterns for harm instructions, drug or weapon synthesis, CSAM, and hate speech.

The system prompt also enforces an explanatory refusal rule: Ophraxx AI is never permitted to say "I can't help with that" without providing a specific reason. Every refusal must name the exact cause and, where possible, offer an alternative.

Conversation Memory

Ophraxx AI maintains per-user conversation sessions within each server. Each session stores up to 20 messages of context and expires after 15 minutes of inactivity. This gives the bot genuine multi-turn memory within a session — users can refer back to earlier parts of a conversation naturally without repeating context. Sessions are stored in memory only and are not persisted to disk or a database.

Live Tool Access

Ophraxx AI can access live external data when the query requires it. The first integrated tool is real-time weather data, pulled from a live weather data provider. When a message is identified as a weather query, the tool dispatcher fetches current conditions — temperature in both Celsius and Fahrenheit, feels-like, humidity, wind speed and direction, visibility, cloud cover, pressure, sunrise and sunset times — and injects this data directly into the system context before the model call. This means the model responds with real, timestamped information rather than disclaiming it cannot access live data.

Usage Limits and Rate Controls

To keep the platform predictable and prevent abuse, Ophraxx AI enforces usage limits per subscription tier across the Ophraxx Web Core:

• OW-F1 (Free): ~500 messages per month
• OW-B12 (Basic): ~1,000 messages per month
• OW-U45 (Ultra): 10,000+ messages per month, priority processing
• OW-P88 (Pro): 50,000+ messages per month, highest priority processing

Limits are checked atomically to prevent race conditions. When a limit is reached, the user receives a clear message specifying which limit was hit and when it resets. Server admins can view live usage statistics — monthly totals, active user counts, and remaining allocation — using the /ophraxx usage slash command.

Hard spam detection is also active: if a user sends too many messages within an 8-second window, the bot declares a spam event, applies a 5-minute soft-block, attempts a 60-second Discord timeout, and sends a detailed embed to the server's mod log channel.

Data and Privacy

Ophraxx AI does not store conversation messages in a persistent database. Session history lives in memory for up to 15 minutes and is discarded when the session expires or the bot restarts.

Usage counts (request counts per user per day, per guild per month) are tracked on our backend to enforce rate limits. User profile records track whether a user has received a welcome message. No message content is written to a long-term database.

Users who want to opt out of having their service interactions used for future model improvement may contact us through our designated support channel.