Some reports stored in Busabase may contain charts rather than data tables. If I ask an Agent to summarize the figures, how should it distinguish values explicitly printed on the chart from values estimated from bar heights or line positions? I want any estimates to remain clearly identifiable in the result.
Question
Closed · AnsweredHow should I ask an Agent to extract values from a chart without overstating precision?
Jen

Ask the Agent to separate printed numbers from geometry estimates, and to keep that label in the output. Charts in Busabase docs or files are pictures of data — an Agent can read labels reliably when they are drawn as text, but bar heights and line positions are estimates unless a table is also present.

*Diagram · Printed vs estimated values in the answer*
Prompt pattern (copy / adapt)
Extract figures from this chart for [topic].
For every number, set source to exactly one of:
- printed — digits visible as text on the chart (axis tick, data label, callout, legend)
- estimated — inferred from bar height, line position, or area; give a range or “≈” and the assumption
- unclear — say what is missing
Output a table: label | value | unit | source | confidence
Do not present estimated values as if they were printed.
If both a printed label and a bar exist and they disagree, report both and flag the conflict.
Prefer the underlying table/CSV in the same document when one exists.How to use it
- Point the Agent at the specific page / image / doc section (avoid “the whole Space”).
- Require the source column in the answer; reject replies that only dump bare numbers.
- For decisions (targets, compliance), use printed values only; treat estimated as directional.
- If the report also has a data table, tell the Agent to extract from the table first and use the chart only as a cross-check.
How to verify
- Every figure in the reply is tagged printed / estimated / unclear.
- Estimated rows use ≈ or a range, not fake four-decimal precision.
- Spot-check two printed labels against the image yourself.
Boundaries
- Chart extraction is not a substitute for the source dataset; precision cannot exceed what the graphic actually encodes.
- Different Agents / models vary in vision quality — the labeling rule matters more than trusting a single run.
- Do not ask for “exact values from the bars” without allowing estimation language; that encourages overstated precision.
If you share whether the chart sits in a Doc image, a PDF, or an attached PNG, we can tighten the prompt with the right “open this node” wording.
Your reply
Sign in to join the discussion. Reading is always open; you only need an account to post.
Sign in