| name | summarize |
| description | Extract structured fields from the bundled filing, write a grounded summary, grade it, and save a report. Usage /summarize |
/summarize
Run the full Extract → Summarize → Grade → Check loop and save a report.
- Extract (extractor agent):
python -m tools.extract > reports/_fields.json
If schema_errors is non-empty, stop and report — do not summarize an invalid extraction.
- Summarize (summarizer agent): write a grounded summary from
reports/_fields.json,
citing the field name behind every figure, and save it to reports/_summary.txt.
- Grade (grader agent):
python -m tools.grade --summary-file reports/_summary.txt
- Gate: if
figure_coverage < 1.0, remove or correct the unsupported_figures and
re-grade; if faithfulness < 0.7, tighten the summary to the extracted fields. Loop at
most 3 times.
- Save to
reports/summary.md:
- the extracted fields and their source spans,
- the final summary with field citations,
- the two scores and the verdict.
Try this to start:
/summarize → summarises the bundled Orion Dynamics Q3 FY2025 release.
To see the faithfulness check bite, make the summarizer state a figure that is not in
reports/_fields.json (e.g. revenue of $2.30 billion) and re-grade: figure_coverage
drops below 1.0 and the loop sends it back.
