NEMOTRON  ·  PERPLEXITY SONAR  ·  GRAPHIFYMD.COM
Nemotron + Perplexity Sonar.
Traversed, not retrieved.
Two stacks. One traversable graph. 83 declared nodes.
NVIDIA Nemotron — the full pipeline from pre-training to Minitron pruning. Perplexity Sonar — five models, real-time web, deep research. Every dependency declared and typed.
nemotron perplexity-sonar read-only
How it's built
Every edge is a decision.
REQUIRES
Hard prerequisite — cannot function without it. Drives agent sequencing.
ENABLES
Unlocks a capability. Not strictly required — optimization paths.
RELATES_TO
Conceptual proximity. Not a dependency. Use for disambiguation.
IMPLEMENTS
Concrete instantiation of an abstraction. Maps architecture.
Confidence defers to null — not wrong, just unreviewed. This is the scaffold your agent runs before you add your domain layer on top.
KRB Benchmark  ·  v0.6.2  ·  open & reproducible
The eval for structured knowledge retrieval.
# System Macro F1 Tokens/q 5-Hop F1
1 CKG (ckg-mcp v0.7.6)
0.488
252 0.786
2 RAG (text-embedding-3-small)
0.123
2,982 0.170
3 GraphRAG (MS global mode)
0.120
3,450 —
CKG F1 improves with hop depth — 0.37 → 0.77 from hop 0 to hop 5. RAG stays flat at ~0.13 regardless of depth. Retrieval has no mechanism for traversing a chain.
↗ danyarm/krb-leaderboard — open & reproducible
Token efficiency
11× fewer tokens.
Context you save, you keep.
269 CKG · tokens/query
vs
2,982 RAG · tokens/query
A grounding pass that costs 269 tokens leaves your context window open for reasoning. Every token saved is a token you can spend on what actually matters to your agent.
Perplexity Sonar  ·  Model Routing  ·  Declared Dependencies
5 Sonar models.
One right traversal.
FAST · CHEAP
sonar
Real-time web. Low latency. Batch retrieval.
PRO · HIGH QUALITY
sonar-pro
Foundation for deep research. More sources, higher quality.
REASONING · NEMOTRON
sonar-reasoning
Multi-step chains. Extended thinking. Nemotron base.
DEEP RESEARCH · BEST
sonar-deep-research
Hours of research in minutes. Structured reports with citations.
Every model's prerequisites declared in the graph. Your agent traverses the right chain — no guessing, no latency from wrong model calls.
1 / 5
Agent Team Orchestration  ·  Liu et al. arXiv:2606.30986
Context degrades.
Traversal doesn't.
In multi-agent pipelines, context efficiency collapses 91% across stages with no model change (Liu et al.). Every agent boundary is a lossy handoff. CKG attacks all three root causes.
TOKEN BURDEN
11×
fewer tokens per query · 269 vs 2,982
HANDOFF COST
0
re-retrievals at agent boundaries
COMPRESSION LOSS
Once
graph compressed offline — never again
Structured context doesn't consume your context window. It opens it.