Invoice AI – LLM-Powered Invoice Extraction (.NET 8)

A production-ready C# .NET 8 project that demonstrates training a base LLM to accurately read, understand, and extract structured data from invoices and receipts of any format.

---

Architecture Overview

InvoiceAI/
├── Controllers/
│   ├── InvoiceController.cs      – POST /api/invoice/extract
│   ├── TrainingController.cs     – POST /api/training/run
│   └── DemoController.cs         – POST /api/demo/run/{id}
├── Services/
│   ├── IInvoiceExtractionService.cs
│   ├── InvoiceExtractionService.cs   – Anthropic Claude API integration
│   └── ModelTrainingService.cs       – Training pipeline orchestration
├── Models/
│   ├── Invoice.cs                – Domain models
│   └── TrainingModels.cs         – Training/request/response DTOs
├── Data/
│   └── SampleInvoiceDataset.cs   – 8 labelled training invoices
├── InvoicePrompts.cs             – System & few-shot prompt engineering
└── Program.cs                    – DI setup + startup auto-training

---

How "Training" Works

This project uses two complementary LLM training techniques:

1. Few-Shot In-Context Learning

Labelled (invoice text → extracted JSON) pairs from the sample dataset are injected into the system prompt at inference time. The LLM learns:

• Your exact JSON output schema
• How to handle different invoice formats (retail, medical, SaaS, construction, etc.)
• Edge cases like deposits, discounts, multi-currency

2. Prompt Engineering / Optimisation

The system prompt is carefully crafted to act as the "weights" of the model. The training service evaluates field accuracy across the dataset and selects the highest-performing prompt variant.

Production Fine-Tuning Path

The POST /api/training/export/jsonl endpoint exports a JSONL file in OpenAI fine-tune format. To submit actual weight-update fine-tuning jobs:

# Export training data
curl https://localhost:5001/api/training/export/jsonl > training.jsonl

# Submit to OpenAI (requires OpenAI API key)
openai api fine_tuning.jobs.create \
  -t training.jsonl \
  -m gpt-4o-mini-2024-07-18

---

Sample Training Dataset

ID	Category	Vendor
train-001	Technology Services	ACME Tech Solutions LLC
train-002	Retail Receipt	Best Buy
train-003	Freelance / Creative	Sarah Chen Design Studio
train-004	SaaS Subscription	Stripe
train-005	Construction / Trades	Reliable Builders Inc.
train-006	Restaurant Receipt	The Golden Fork
train-007	Medical / Healthcare	City Medical Center
train-008	Shipping / Logistics	FedEx Freight

---

Quick Start

Prerequisites

• .NET 8 SDK
• An Anthropic API key

1. Configure your API key

Option A – appsettings.json:

{
  "Anthropic": { "ApiKey": "sk-ant-..." }
}

Option B – Environment variable (recommended):

export ANTHROPIC_API_KEY=sk-ant-...

2. Run the project

cd InvoiceAI
dotnet run

The app starts on https://localhost:5001. Swagger UI loads at the root /.

3. Try the API

Run training pipeline:

curl -X POST https://localhost:5001/api/training/run

Extract a sample invoice (fine-tuned):

curl -X POST https://localhost:5001/api/demo/run/train-001

Compare base vs fine-tuned:

curl -X POST https://localhost:5001/api/demo/compare/train-002

Extract your own invoice:

curl -X POST https://localhost:5001/api/invoice/extract \
  -H "Content-Type: application/json" \
  -d '{
    "invoiceText": "ACME Corp\nInvoice #123\nDate: 2024-01-15\n\nWeb Development  1  $5000\n\nTotal: $5000",
    "useFineTunedModel": true
  }'

---

API Reference

POST /api/invoice/extract

Extract structured data from any invoice text.

Request:

{
  "invoiceText": "...",
  "useFineTunedModel": true,
  "sourceFileName": "invoice.txt"
}

Response:

{
  "success": true,
  "invoice": {
    "vendorName": "ACME Corp",
    "invoiceNumber": "INV-123",
    "invoiceDate": "2024-01-15T00:00:00",
    "totalAmount": 5000.00,
    "currency": "USD",
    "lineItems": [...],
    "metadata": {
      "confidenceScore": 0.97,
      "wasFineTuned": true,
      "modelVersion": "claude-opus-4-5"
    }
  },
  "processingTimeMs": 1423
}

POST /api/training/run

Runs the full training pipeline and returns metrics.

GET /api/training/stats

Returns dataset statistics (example counts, categories, coverage).

GET /api/training/export/jsonl

Downloads the training dataset in OpenAI JSONL format.

POST /api/demo/run/{sampleId}

Runs extraction on a built-in sample invoice. IDs: train-001 to train-008.

POST /api/demo/compare/{sampleId}

Runs both base and fine-tuned extractions side-by-side for comparison.

---

Extending the Training Dataset

Add a new TrainingExample to SampleInvoiceDataset.cs:

public static TrainingExample GetMyNewInvoice() => new()
{
    Id = "train-009",
    Category = "utilities",
    RawText = """
        POWER COMPANY
        Account: 123456
        Amount Due: $145.50
        Due: March 1, 2024
        """,
    GroundTruth = new Invoice
    {
        VendorName = "POWER COMPANY",
        TotalAmount = 145.50m,
        InvoiceDate = new DateTime(2024, 3, 1),
        // ...
    }
};

Then register it in GetAllExamples():

public static List<TrainingExample> GetAllExamples() => new()
{
    // existing examples...
    GetMyNewInvoice()
};

---

Key Technologies

Technology	Purpose
.NET 8	Runtime & web framework
ASP.NET Core	REST API
Anthropic Claude API	Base LLM for extraction
Newtonsoft.Json	JSON parsing
Swashbuckle	Swagger UI
Microsoft.Extensions.ML	ML.NET integration hooks