Food Bills pipeline
A raw → wiki → reports knowledge pipeline that turns grocery receipts into an auditable price-and-spend knowledge base, driven by a custom /add-bill agent skill.
Problem
Useful grocery insights appear across receipts, but spreadsheets usually discard the source. I wanted an agent to handle transcription while keeping every price, total and conclusion traceable to the exact receipt it came from.
What I built
A plain-file pipeline: immutable receipts in raw/, curated bill and product notes in wiki/, and disposable reports plus a self-contained dashboard in reports/. A custom /add-bill skill transcribes new receipts, resolves product aliases, updates price histories and regenerates affected outputs.
The hard part
The hard part was making LLM-authored data trustworthy without pretending the model is perfect. Flow is strictly one-way: corrections happen in the wiki, never in a report. Digital PDFs require cent-perfect reconciliation; uncertain photo lines are marked low-confidence and excluded from price comparisons. Unreadable values stay unreadable—nothing is invented to make a total work.
Receipt text is also treated as untrusted data, never as instructions. Together, source traceability and confidence gating make errors visible and recoverable.
Results
The pipeline currently covers 12 receipts, three stores and 43 canonical products. Every report figure links back through a wiki note to a real receipt, while one repeatable command handles the tedious ingestion work.