Tokenomics Framework
Why Tokenomics Matters
Standard IR benchmarks measure quality only. In production LLM systems, token cost is a first-class constraint:
- Enterprise AI budgets are token-denominated
- Latency scales with token count
- Context window limits bound what can be retrieved
- Cost per query determines whether a system is deployable at scale
Token Profiles
RAG
| Metric | Value |
|---|---|
| Tokens per query | 3,000--5,000 |
| Noise ratio | High (off-topic chunk content) |
| Build cost | Embed all chunks |
GraphRAG
| Metric | Value |
|---|---|
| Tokens per query | 2,000--8,000 (high variance) |
| Noise ratio | Medium (compressed summaries) |
| Build cost | HIGH (full text extraction) |
CKG
| Metric | Value |
|---|---|
| Tokens per query | 150--400 |
| Noise ratio | Near-zero (exact subgraph) |
| Build cost | ZERO (CSV already exists) |
Pricing Model
All costs computed at Claude Sonnet 4.6 pricing:
- Input: $3 per 1M tokens
- Output: $15 per 1M tokens
The Structure Premium Hypothesis
The token efficiency gap between CKG and RAG is proportional to the structural richness of the domain's DAG:
\[\text{dag\_richness}(d) = \frac{\text{edges}}{\text{concepts}} \times \text{mean\_indegree} \times \frac{1}{\text{orphan\_rate}}\]
If structure_premium correlates with dag_richness (\(r > 0.7\)), this validates the core claim: explicit structure is the source of the efficiency gain.