Read the repository.
A content pass reads tracked files and extracts declarations, imports, calls and comments. Compact descriptions retain distinctive facts. A lexical index makes the source searchable.
Read the indexing codeUnder the surface
The source is
the explanation.
System One Search pairs a repository index with small, typed judgments. Code finds the candidates; Jev, TypeSafe’s decision model, checks their relevance.
The flow chart is built entirely with TypeSafe decisions and code. TypeSafe judges which files belong, which source passages explain them, and which references matter. Code assembles those decisions into the chart. No generative LLM writes the explanation.
Find the candidates.
Paths, facts and source text propose relevant files. A static reference graph records connections between them. Retrieval is computed in code, before asking the model to judge.
Read the lexical indexCheck the evidence.
TypeSafe evaluates relevance and checks candidates against source evidence. Questions stay constant; repository content goes in the request state. The model returns typed judgments, not generated prose.
Read the actual questionsConnect the source.
For a “how” question, the search expands through references, judges evidence and assembles a flow. The chart’s titles are identifiers, summaries are source comments, and arrows represent recorded references.
Read the flow builderEvidence you can inspect
A starting point for understanding.
The live search shows events reported by the engine. No simulated thoughts or progress. Open a result to read its source, or take the same completed run into the playground to inspect the search steps and repository map.
A relevance judgment is not proof. Search depth, caps, approximate extractors and dynamic code can leave gaps. A source flow is a bounded static view, not a runtime trace. Coverage checks record examined paths and remaining reference neighbors; they do not prove complete understanding.