Debugging
When a graph looks wrong, inspect diagnostics before changing the extractor.
Inspect these surfaces first
result.diagnostics.warningsresult.diagnostics.errorsresult.diagnostics.timingsresult.diagnostics.modelUsageresult.artifacts.report
Useful questions
- Did the input items have stable IDs and meaningful titles?
- Did the local path scan include files you did not expect?
- Are semantic edges doing work that should actually be deterministic?
- Did the graph contain nodes but no meaningful cross-source relationships?
Fast debugging loop
- Start from
examples:from-textsorexamples:llm-mock. - Inspect
reportandgraph.stats. - Run
query.path(...)andquery.neighbors(...)against the nodes you care about. - Only then modify extraction or semantic logic.

