Specify the document task
Describe whether the model should locate tables, extract cells, connect notes, or answer questions from supplied reports. Keep the project focused on document understanding rather than assumptions about investment usefulness. Identify the relevant report types, accounting periods, languages, and public or reviewed private sources.
Preserve period and unit context
A value may be expressed in thousands, millions, percentages, or per-share terms. Capture column dates, currencies, scaling notes, and whether figures describe a quarter or cumulative period. Include those details in expected outputs so a text match does not incorrectly count a numerically misinterpreted answer as correct.
Handle comparisons and revisions
Reports may present current and prior periods, revised figures, or different organizational segments. Record which column and version supports each label. If a task compares periods, require the model to use the designated source context rather than silently substituting a later restated value from another document.
Review evidence-based outputs
Attach page and table references to extracted values and answers. Include no-answer cases when a requested measure is absent. Send Fiund representative report pages and an output schema, together with source-review requirements and exclusions, so any proposed sourcing plan can be evaluated against the actual extraction or retrieval task. In a pilot, ask for a value shown under a heading such as amounts in millions and compare it with the normalized output. Check the sign, period, currency, and scale separately. This exercise catches answers that match visible digits but represent the wrong quantity once the document context is applied.
Review checklist
- Define extraction versus question-answering tasks
- Record currency, scale, and reporting period
- Track report versions and restatements
- Require page and table evidence