Carrefour Receipt OCR Workflow with Mistral, Gemini & Google Sheets

Turn receipt images into structured data—with confidence checks and human approval

This Carrefour Receipt OCR n8n workflow receives receipt images via Webhook or Telegram, runs OCR using Mistral with word-level confidence scoring, routes low-confidence documents to Gmail approval, then extracts fields and line items with Google Gemini and upserts them into Google Sheets—while logging every step to Postgres.

What this workflow does

• Ingests a receipt from a POST webhook request (preferred for accuracy) or an incoming Telegram message attachment (downloads the image).
• Converts the file to a base64 data URL and sends it to the Mistral OCR API for text extraction with word-level confidence.
• Calculates confidence metrics (minimum/average) and flags low-confidence words; stores execution metadata and inserts an audit row into the Postgres document_processing_log table.
• If OCR confidence is below the threshold, sends a Gmail “send and wait” approval email containing the OCR text and flagged words, then updates the log as approved or rejected based on the response.
• If approved (or if no review is needed), uses Google Gemini with a structured output schema to extract receipt header fields and item lines from the OCR text.
• Updates Google Sheets: splits extracted items into individual rows and appends or updates them in the receipts sheet.
• Finalizes and notifies: updates Postgres status (including “inserted”) and emails a success or rejection notification with the original document attached.

Use cases

• Carrefour receipt capture for expense tracking with an audit trail.
• Human-in-the-loop review when OCR confidence is uncertain.
• SaaS operations where receipt data must be reliably stored in Google Sheets and logged in Postgres.

Technical details

• n8n nodes/tools: webhook, telegram (download attachment), code (confidence scoring + data handling), Mistral OCR API, Gemini (structured extraction), Google Sheets (append/update), Gmail (send and wait + notifications), Postgres (document_processing_log), and sticky note for workflow guidance.