Skip to main content

Tips and Best Practices

Choosing the Right Model​

For cost-conscious users:

  • grok-code-fast-1 — Fastest, lowest cost, surprisingly good quality
  • GPT-4o mini — Affordable, decent quality
  • Gemini Flash 1.5 — Fast, low cost, huge context window

For best quality:

  • GPT-4o — Best vision support, excellent reasoning
  • Claude Opus 4 — Top-tier quality, 200k context
  • Grok 4 — Latest from xAI, strong UE5 knowledge

For vision (screenshots):

  • GPT-4o — Industry standard, reliable
  • Grok Vision — Good balance of speed and quality
  • Claude 3.5 Sonnet — Excellent vision + reasoning
  • Gemini Pro Vision — Solid option, low cost

For huge Blueprints/context:

  • Claude Opus 4 — 200k tokens (fits massive Blueprints)
  • Gemini Pro 1.5 — 1M+ tokens (can handle entire projects)

Asking Effective Questions​

Good questions are specific and action-oriented:

✅ Good:

  • "How do I spawn an actor at the player's location in C++?"
  • "Analyze this Blueprint for unused variables"
  • "What's the performance impact of using Tick vs. Timer?"
  • "Generate a health component with TakeDamage and Heal functions"

❌ Less effective:

  • "Help" (too vague—help with what?)
  • "Fix my code" (no context—what code? what's wrong?)
  • "Why doesn't it work?" (need error message, logs, or Blueprint context)
  • "Make it better" (better how? performance? readability?)

Rule of thumb: If you were asking a colleague, would they understand the question without seeing your screen? If not, add context (screenshot, Blueprint context, log lines).

Providing Context​

The more context you provide, the better the AI can help:

For compilation errors:

  • Enable "Include Last Log Lines in Query"
  • Settings > Interface > Include Last Log Lines in Query (check ON)
  • Set Last Log Lines Count to 20-50
  • Ask: "What's causing this error?"

For visual issues:

  • Capture a screenshot (Blueprint toolbar Screenshot button)
  • Use a vision-capable model (GPT-4o, Grok Vision, Claude 3.5+)
  • Ask: "What's wrong with this node layout?"

For logic errors:

  • Send Blueprint context (toolbar or right-click)
  • Switch to Deep Scan mode
  • Ask: "Why doesn't this event fire?"

For asset questions:

  • Right-click the asset > Describe to Orion
  • Ask follow-up questions with context

Managing Token Costs​

Tokens = API fees. Here's how to keep costs low:

Use fast/cheap models for simple questions:

  • grok-code-fast-1 costs ~1/10th of GPT-4o
  • GPT-4o mini costs ~1/5th of GPT-4o
  • Save expensive models for complex analysis

Disable conversation context when not needed:

  • Settings > Interface > Use Conversation Context (uncheck)
  • Each query becomes independent (no history sent)
  • Big savings if you don't need multi-turn discussions

Use "Graph Flow" instead of "Full Graph":

  • Graph Flow analyzes only execution nodes (smaller payload)
  • Full Graph includes all data links (larger payload)
  • Graph Flow can be 30-50% fewer tokens

Select specific nodes for large Blueprints:

  • Don't send 200-node Blueprint for a 10-node question
  • Select the relevant section, Send > Selection
  • Can reduce payload by 80-90%

Clear server history periodically:

  • xAI conversation history accumulates over time
  • Settings > "Clear server history" button
  • Prevents old context from inflating future queries

Set Max Output Tokens:

  • Settings > AI Configuration > Max Output Tokens
  • Limit response length if you don't need essays
  • Example: Set to 2000 for concise answers (default 0 = unlimited)

Security Best Practices​

API key safety:

  • Never share your API keys with others
  • Never commit keys to version control (Orion stores them in user profile, not project files)
  • Set key expiry (30 days recommended) so old keys auto-expire
  • Validate keys regularly (Settings > Validate button)

Query review before sending:

  • Orion never sends anything without your explicit action (clicking Send, right-click option, toolbar button)
  • Review what you're sending, especially if your project has proprietary logic
  • Use "Describe to Orion" for metadata-only analysis (no full code/graph sent)

Context injection settings:

  • "Include Last Log Lines" — only enable when troubleshooting (logs may contain sensitive data)
  • Project name and engine version are always included to help the AI provide relevant answers

API provider privacy policies:

  • Review xAI, OpenAI, Anthropic, Google privacy policies
  • Understand what data is retained (most providers retain queries for 30 days)
  • For maximum privacy: Use EDA export (local AI) instead of cloud APIs