TrackitOcr - An Invoice Parser (Elixir + Mistral Document AI)

This project demonstrates how to use Mistral’s Document AI QnA API to extract structured data from uploaded documents — specifically invoices — using Elixir.

It serves as a learning and experimentation project to try out:

• How to interact with Mistral’s Document AI API.
• How to use structured output in chat completions.

---

🧰 Tech Stack

• Language: Elixir ~> 1.18
• HTTP client: Req
• AI API: Mistral Document AI QnA

---

⚙️ How It Works

1. File Upload

   The document (e.g. an invoice PDF) is first uploaded to Mistral’s AI blob store via their /files endpoint.

   This returns a document_url which can be referenced later.

2. Chat Completion Request

   The document_url is then passed to the chat/completions API along with a short instruction message and a JSON schema defining the structured response format.

   Example schema fields:

   • amount_to_pay_cents
   • invoice_date
   • invoice_number
   • currency
   • reason_for_payment
   • issuer

3. Structured Output

   The API returns a JSON object matching the schema

---

🚀 Running Locally

Prerequisites

• Elixir ~> 1.18
• mix toolchain
• A valid Mistral API key

Setup

git clone <git@github.com:acramatte/trackit-ocr.git> cd trackit-ocr mix deps.get

Environment variables

Export your API key:

export MISTRAL_API_KEY="your_api_key_here"

Run

You can run the module or script directly via iex:

iex -S mix

Then call the function passing it the path to the file to upload:

Trackit.processPDF("../2025-11-04_My_Invoice_080000463123.pdf")