Context Priority Inversion

Issue: Low-Priority Content Displaces High-Priority Content

Frequency: Common

Symptoms

• Verbose tool outputs consume context
• Irrelevant RAG results included
• Important instructions at risk
• Recent low-value content kept over old high-value
• System prompt gets squeezed

Root Cause Context assembly often doesn’t consider content priority. A 5,000-token tool output is included fully while critical instructions get truncated. Without priority-based assembly and truncation, context fills with low-value content, displacing what matters most.

Example

Context budget: 8K tokens

Priorities should be:
1. System prompt (critical)
2. User preferences (high)  
3. Current query (high)
4. Relevant context (medium)
5. Tool outputs (low)

Actual assembly (no priority):
- System prompt: 1K tokens
- Tool output (full JSON dump): 6K tokens
- Current query: 500 tokens
- [NO ROOM] User preferences: truncated
- [NO ROOM] Relevant history: truncated

Result: 75% of context is verbose tool output
        Critical preferences lost

Contributing Factors

• No content prioritization
• Tools return verbose output
• RAG returns full documents
• FIFO assembly without scoring
• No content compression
• Fixed allocation per content type

Eval Recipes

Test Cases

Metrics

Mitigation Strategies

Prevention

1. Priority tiers: Define strict priority levels
2. Reserved budgets: Guarantee space for critical content
3. Tool output limits: Cap/summarize tool responses
4. RAG filtering: Only include relevant chunks
5. Dynamic allocation: Adjust based on content value
6. Compression by tier: Compress low-priority first

Priority Framework

Tier 1 (Protected): System prompt, safety rules
Tier 2 (Reserved): User preferences, key facts
Tier 3 (Allocated): Current query, recent turns
Tier 4 (Best effort): Tool outputs, RAG context
Tier 5 (Compressed): History, verbose content

Production Signals

Key Metrics

Alerts

References

• LangChain: Context Management
• Context Window Optimization