Compute Units

Compute Units (CU) represent the computational budget allocated to your research queries, directly influencing the depth and breadth of Caesar’s reasoning process. Each unit roughly corresponds to one minute of processing time and enables additional agentic loops for information gathering, validation, and synthesis.

Understanding compute units

Compute Units control the number of reasoning steps Caesar performs during research. Think of them as iterative cycles where Caesar:

1. Gathers information from diverse sources
2. Analyzes and synthesizes findings
3. Identifies gaps in the current understanding
4. Validates conclusions against additional evidence
5. Refines the response based on accumulated knowledge

Impact on research quality

The relationship between compute units and research quality is nuanced. More compute doesn’t always mean better resultsit depends on the nature of your query.

Key observations

1. CU 1 (19.95%): Single-pass reasoning with limited context gathering
2. CU 2-3 (39.83% - 53.85%): Substantial improvement as Caesar can validate and refine initial findings
3. CU 10 (55.87%): Marginal gains beyond CU 3, showing diminishing returns for most queries

The dramatic improvement from CU 1 to CU 2-3 occurs because Caesar gains the ability to:

• Cross-reference initial findings
• Fill knowledge gaps identified in the first pass
• Validate conclusions against contradictory evidence
• Synthesize a more coherent narrative

Choosing the right compute budget

Query complexity matrix

Examples by domain

1
// Earnings summary (CU: 2)
2
{
3
  "query": "Apple Q3 2024 earnings highlights",
4
  "compute_units": 2
5
}
6
7
// Market comparison (CU: 5)
8
{
9
  "query": "Comparative analysis of FAANG stocks YTD performance with sector rotation implications",
10
  "compute_units": 5
11
}

The reasoning cascade

Understanding how Caesar utilizes compute units helps optimize your usage:

CU 1: Foundation

• Initial query interpretation
• Primary source retrieval
• Basic synthesis
• Limited validation

CU 2-3: Validation

• Cross-source verification
• Contradiction resolution
• Gap identification and filling
• Enhanced coherence

CU 4-6: Expansion

• Peripheral context gathering
• Expert perspective integration
• Nuanced analysis
• Robust fact-checking

CU 7-10: Exhaustive

• Edge case exploration
• Minority viewpoint inclusion
• Deep technical validation
• Comprehensive citation network

Optimization strategies

Technical implementation

When you specify compute units in your request:

$
curl -X POST https://api.caesar.xyz/research \
>
  -H "Authorization: Bearer YOUR_API_KEY" \
>
  -H "Content-Type: application/json" \
>
  -d '{
>
    "query": "Your research question",
>
    "compute_units": 5
>
  }'

Caesar internally:

1. Allocates reasoning cycles proportional to the CU value
2. Expands search breadth and depth parameters
3. Enables additional validation passes
4. Increases synthesis complexity thresholds

Performance characteristics

Key takeaways

1. CU 2-3 represents the sweet spot for most research queries, providing substantial multi-step reasoning without excessive computation
2. The jump from CU 1 to CU 2 is most significant, enabling validation and gap-filling capabilities
3. Higher CU values (>5) show diminishing returns except for genuinely complex, quantitative research
4. Match CU to query complexity, not query importanceeven critical queries may only need CU 2-3 if they’re straightforward
5. Consider time-to-insight tradeoffssometimes multiple focused lower-CU queries beat a single CU 10 query