Context graph vs RAG: the numbers on tokens, cost, and noise
How much a context graph saves over RAG, in plain figures: about 85% fewer tokens sent to the model, lower cost per answer, and less noise.
RAG and a context graph both feed a model context, but they send very different amounts of it. This piece puts numbers on the difference. The short version: Suda passes about 85% less text to the model than typical retrieval, which lowers cost per answer and cuts the noise that causes wrong answers.
The headline number
A context graph sends only the context an answer needs, not the top-k passages that merely look similar. In practice that is about 85% fewer tokens reaching the model per query.
- RAG: retrieve the most similar chunks, stuff them into the prompt, let the model sort it out.
- Context graph: resolve what is actually relevant and current, send that slice.
What that means per answer
The table below uses a representative internal question (“what is our current refund window, and what changed it”) to make the difference concrete. Token counts are illustrative of the pattern, not a benchmark of any one model.
| RAG (top-k retrieval) | Context graph (Suda) | |
|---|---|---|
| Context sent to model | ~12,000 tokens | ~1,800 tokens |
| Reduction | baseline | ~85% fewer |
| Conflicting facts in context | common (several chunks) | resolved to the current one |
| Stale facts in context | possible | retired automatically |
| Sources traceable | ranked passages | connected, cited nodes |
| Relative cost per answer | baseline | lower, in proportion to tokens |
Why fewer tokens also means better answers
Cost is the obvious win, but the quality win matters more. When you dump 12,000 tokens of similar-looking passages into a prompt, some of them contradict each other or are out of date, and the model can anchor on the wrong one. Sending a smaller, resolved slice removes that failure mode: fewer tokens, less contradiction, fewer wrong answers.
- Lower cost: fewer tokens in equals a smaller bill per query, at scale.
- Less noise: the model is not asked to referee stale or conflicting chunks.
- Traceability: answers come from connected, current nodes you can follow.
How Suda gets there
Suda builds a context graph from your tools, models how facts connect and which is current, retires stale facts, resolves conflicts, and hands an agent only what is needed, over MCP.
npx suda connectFor the architecture behind these numbers, read context graph vs RAG. For the concept, read what is a company brain.