Skip to main content

​
Start from an outcome

These recipes are prompt shapes, not guaranteed results. Replace file, table, column, unit, and date names with values confirmed by /files and /schema.

​
Inspect before calculating

Outcome: confirm that the source Fella will query matches your assumptions.
Check inferred numeric types, date formats, nulls, units, capitalization, skipped files, preamble removal, and dropped total-row notes. Fix or clarify bad source assumptions before asking for a total.

​
Compare two periods

Outcome: get two totals and a reproducible delta.
Inspect the date boundaries, category filter, refund handling, denominator, and units in the working fold.

​
Rank categories

Outcome: find the largest contributors rather than list raw rows.
If you mean completed workouts only, say so and name the status value. Verification cannot infer a missing business rule.

​
Join two files

Outcome: combine records through an explicit key.
Inspect that both tables appear in the SQL, the join key is correct, and filters run before aggregation. If keys have different types or spelling, normalize the source or state the intended mapping.

​
Reconcile a document and a table

Outcome: use a stated target from text with observed values from rows.
Text claims appear as document evidence; calculations should still be grounded in tool output. A scanned PDF needs OCR before this recipe works.

​
Find and summarize document facts

Outcome: locate relevant passages without an embeddings index.
Fella uses direct regex search and document reads. Ask for file names or quoted passages when source traceability matters.

​
Create a compact chart

Outcome: visualize a small aggregated result.
Charts support bar and line data with at most 12 categories and 2 series. Aggregate first instead of requesting thousands of points. The chart is derived from a read-only query result, so its labels and values are recorded with the same evidence as the calculation behind it.

​
Query directly

Outcome: bypass model interpretation when you know the table and SQLite query.
Direct SQL remains read-only and prints a system result; unlike an agent tool call, it does not create an evidence fold or run post-answer verification. The shipped backend is SQLite, so use SQLite syntax. A DuckDB-only function will not work in a normal release build.

​
Ask an unanswerable question safely

Outcome: separate recorded facts from a forecast.
The prompt improves the chance of a proper refusal, but policy adherence remains model-dependent. Review the evidence whenever Fella supplies a number.

​
Next steps

Troubleshoot a recipe

Diagnose source, provider, tool, and interpretation failures.

Inspect evidence

Review source selection, query logic, rows, and checks.