The README tells you to "run both on the same question and compare what each one retrieved and cited", and then leaves you to do it by hand: ask, read, switch method, ask again, remember what the first one said.
Local and global reach different things. Local walks out from entities it ranked; global reads community summaries and never sees an entity that no community claims. Which one suits a question is the decision somebody has to make about their own data, and the tab has everything needed to show it.
What to build
One question, both runs, read together. The evidence sets diffed: what only local retrieved, what only global retrieved, what both did, and which of those each answer actually cited.
- The two runs are independent, so they can go at once.
SearchRun already carries the full context and the citations; countUsage and citedShortIds in src/core/search/highlight.ts already separate cited from retrieved.- A trace file holds a list of runs (
runs: SearchRun[]), so a comparison is savable and re-openable with no extra format.
Care about
- The cost doubles. Say so before it is spent, the way the global plan line does.
- Do not imply one method won. The point is what each one reached, not a score.
Done when
Local and global answers to one question are read together, with what only one of them retrieved marked, and the pair saves and re-opens as one trace.
Roadmap: M11.
The README tells you to "run both on the same question and compare what each one retrieved and cited", and then leaves you to do it by hand: ask, read, switch method, ask again, remember what the first one said.
Local and global reach different things. Local walks out from entities it ranked; global reads community summaries and never sees an entity that no community claims. Which one suits a question is the decision somebody has to make about their own data, and the tab has everything needed to show it.
What to build
One question, both runs, read together. The evidence sets diffed: what only local retrieved, what only global retrieved, what both did, and which of those each answer actually cited.
SearchRunalready carries the full context and the citations;countUsageandcitedShortIdsinsrc/core/search/highlight.tsalready separate cited from retrieved.runs: SearchRun[]), so a comparison is savable and re-openable with no extra format.Care about
Done when
Local and global answers to one question are read together, with what only one of them retrieved marked, and the pair saves and re-opens as one trace.
Roadmap: M11.