Extract bank-statement pages in fresh Gemini chats and check the rows
At a glance
An HN contributor narrowed Gemini extraction to individual statement pages after plausible multi-page CSV output failed verification. This workflow is listed under Money & Bills, Document Extraction, Verification.
SourceBuilding a tool for these workflows? See weekly rates.
What they did
1) Obtain the older bank-statement PDF, Gemini and a PDF splitting tool such as pdftk. The source needed CSV for transactions no longer available through the bank's direct export. 2) Request transaction extraction; if a long document exceeds output limits, split it, but do not accept plausible-looking chunks on spot checks alone. 3) Compare consecutive balances with transaction amounts and inspect original rows at every split boundary for missing, duplicated or misaligned entries. The author's four-page batches failed this check. 4) Extract each page in a fresh chat with a fresh prompt and retry pages that fail balance checks. Keep the PDF as the reference when reviewing the resulting data.
How it went
Fresh contexts improved balance checks, but the author could not establish a time saving.
Worth knowing
Earlier output silently altered rows and balances; internal arithmetic consistency did not prove faithful transcription.
