On a question about one module, a Keystone slice gave the AI every fact it needed with 77% fewer input tokens than the whole graph: about 22,600 instead of 97,900. This page shows how we measured it and where a slice is not enough.
What a slice is
get_architecture can return the whole graph or a part of it (see AI assistant):
- module: the components of one module of a container in full, their neighbours as outlines;
- element + neighbors: one element in full, every relationship touching it and the elements on the other side;
- element: one element with everything inside it, for example a container or a database.
Outlines carry names and kinds but no operations or columns, so the AI reads an outlined element on its own before it answers about its insides.
How we measured
- Project: an anonymized copy of a real 281-element architecture with 433 relationships; names and contracts replaced, the structure kept.
- Model: Claude Haiku. Trials: 120, each question asked 3 times in two modes: with the slice only, and with up to two rounds of reading more.
- Questions: the boundaries of a container, the members and API of a module, the direct and two-hop callers of a component, a table with an end-to-end scenario, and a control question with no answer in the data.
- Scoring: the answer's facts against a list prepared from the full graph before any model call; tokens as the model provider counted them, system prompt and question included.
Results
| Question | Whole graph | Best slice | Slice | Facts found |
|---|---|---|---|---|
| Members and API of a module | 97,905 tokens | 22,567 tokens (−77%) | module | 100%, same as the whole graph |
| Callers of a component, two hops | 97,929 tokens | 22,591 tokens (−77%) | module | 100% |
| Relationships crossing a container | 97,907 tokens | 72,614 tokens (−26%) | element + neighbors | 99% |
| A table and an end-to-end scenario | 97,922 tokens | — | whole graph | 100% |
The whole graph answered every question. In the control question no variant invented an answer.
Where a slice is not enough
- A slice that does not fit the question costs more. For the table-and-scenario question, starting from a module and reading on took about 144,000 tokens, and from a container about 247,000: up to 2.5 times the whole graph.
- One element with its neighbours is cheap but shallow: about 8,200 tokens, yet it covers one hop. Two-hop callers reached 43% of the facts, and 81% after reading more.
- Questions about the whole system — every boundary, a scenario across services — need the whole graph.
That is why Keystone's instructions recommend that the AI treat the summary as a map, pick the slice that fits the task, and read what is missing before it answers.
Limits of this study
One project, one model, three repetitions per question. Other models may read on more or less willingly, and other architectures may give other numbers. We will repeat the measurement as Keystone and the models change.