language:
- en license: mit library_name: transformers tags:
- finance
- entity-extraction
- ner
- phi-3
- production
- indian-banking base_model: microsoft/Phi-3-mini-4k-instruct pipeline_tag: text-generation
Finance Entity Extractor (FinEE) v1.0
Extract structured financial data from Indian banking messages.
94.5% field accuracy. <1ms latency. Zero setup.
⚡ Install & Run in 10 Seconds
pip install finee
from finee import extract
r = extract("Rs.2500 debited from A/c XX3545 to swiggy@ybl on 28-12-2025")
print(r.amount) # 2500.0
print(r.merchant) # "Swiggy"
print(r.category) # "food"
No model download. No API keys. Works offline.
📋 Output Schema Contract
Every extraction returns this guaranteed JSON structure:
{
"amount": 2500.0, // float - Always numeric
"currency": "INR", // string - ISO 4217
"type": "debit", // "debit" | "credit"
"account": "3545", // string - Last 4 digits
"date": "28-12-2025", // string - DD-MM-YYYY
"reference": "534567891234",// string - UPI/NEFT ref
"merchant": "Swiggy", // string - Normalized name
"category": "food", // string - food|shopping|transport|...
"vpa": "swiggy@ybl", // string - Raw VPA
"confidence": 0.95, // float - 0.0 to 1.0
"confidence_level": "HIGH" // "LOW" | "MEDIUM" | "HIGH"
}
🔬 Verify Accuracy Yourself
Don't trust "99% accuracy" claims. Run the benchmark:
# Clone and test
git clone https://github.com/Ranjitbehera0034/Finance-Entity-Extractor.git
cd Finance-Entity-Extractor
pip install finee
# Run benchmark
python benchmark.py --all
Test on YOUR data:
python benchmark.py --file your_transactions.jsonl
💀 Torture Test (Edge Cases)
Real bank SMS is messy. Here's how FinEE handles the chaos:
| Edge Case | Input | Result |
|---|---|---|
| Missing spaces | Rs.500.00debited from A/c1234 |
✅ amount=500.0 |
| Weird formatting | Rs 2,500/-debited dt:28/12/25 |
✅ amount=2500.0 |
| Mixed case | RS. 1500 DEBITED from ACCT |
✅ amount=1500.0, type=debit |
| Unicode symbols | ₹2,500 debited from •••• 3545 |
✅ amount=2500.0 |
| Multiple amounts | Rs.500 debited. Bal: Rs.15,000 |
✅ amount=500.0 (first) |
| Truncated SMS | Rs.2500 debited from A/c...3545 to swi... |
✅ amount=2500.0 |
| Extra noise | ALERT! Dear Customer, Rs.500 debited... Ignore if done by you. |
✅ amount=500.0 |
Run torture tests:
python benchmark.py --torture
🏦 Supported Banks
| Bank | Debit | Credit | UPI | NEFT/IMPS |
|---|---|---|---|---|
| HDFC | ✅ | ✅ | ✅ | ✅ |
| ICICI | ✅ | ✅ | ✅ | ✅ |
| SBI | ✅ | ✅ | ✅ | ✅ |
| Axis | ✅ | ✅ | ✅ | ✅ |
| Kotak | ✅ | ✅ | ✅ | ✅ |
🏗️ Architecture
Input Text
│
▼
┌─────────────────────────────────────────────────────────────┐
│ TIER 0: Hash Cache (<1ms if seen before) │
└─────────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────┐
│ TIER 1: Regex Engine (50+ battle-tested patterns) │
│ Extract: amount, date, reference, account, vpa, type │
└─────────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────┐
│ TIER 2: Rule-Based Mapping (200+ VPA → merchant) │
│ Map: vpa → merchant, merchant → category │
└─────────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────┐
│ TIER 3: LLM (Optional, for edge cases) │
│ Targeted prompts for: merchant, category only │
└─────────────────────────────────────────────────────────────┘
│
▼
ExtractionResult (Guaranteed Schema)
📊 Benchmark Results
| Metric | Value |
|---|---|
| Field Accuracy | 94.5% |
| Latency (Regex) | <1ms |
| Latency (LLM) | ~50ms |
| Throughput | 50,000+ msg/sec |
| Banks Tested | 5 (HDFC, ICICI, SBI, Axis, Kotak) |
💻 CLI Usage
# Extract from text
finee extract "Rs.500 debited from A/c 1234"
# Show version
finee --version
# Check available backends
finee backends
📁 Repository Structure
Finance-Entity-Extractor/
├── src/finee/ # Core package (16 modules)
│ ├── extractor.py # Pipeline orchestrator
│ ├── regex_engine.py # 50+ regex patterns
│ ├── merchants.py # 200+ VPA mappings
│ └── backends/ # MLX, PyTorch, GGUF
├── tests/ # 88 unit tests
├── examples/ # Colab notebook
├── experiments/ # Research notebooks
├── benchmark.py # ⭐ Verify accuracy yourself
├── pyproject.toml
└── README.md
🤝 Contributing
git clone https://github.com/Ranjitbehera0034/Finance-Entity-Extractor.git
cd Finance-Entity-Extractor
pip install -e ".[dev]"
pytest tests/
📄 License
MIT License - see LICENSE
Made with ❤️ by Ranjit Behera
PyPI · GitHub · Hugging Face
Metadata
Release files for finee 1.0.3
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| finee-1.0.3.tar.gz | 29.4 kB | Without provenance | Details for finee-1.0.3.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| finee-1.0.3-py3-none-any.whl | Python 3 | none | any | Details for finee-1.0.3-py3-none-any.whl |
Total release size: 66.2 kB
Release files / finee-1.0.3.tar.gz
| Download URL | finee-1.0.3.tar.gz |
|---|---|
| Size | 29.4 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
314d6b0a34aef7524fbe990d8d81d0b36cb8c6294c0eb374a1d8e4590e025e5e
|
|
BLAKE2b-256 checksum How to use checksums |
34d018e6e5e0b0fe134c1435eebeb9f5122b309cf709797dd94c791c4fdf30fb
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.9.6
|
Release files / finee-1.0.3-py3-none-any.whl
| Download URL | finee-1.0.3-py3-none-any.whl |
|---|---|
| Size | 36.8 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
679df5028567e484cd8d357102a34ddca6274ad1be6528a926fb9fae9356be9d
|
|
BLAKE2b-256 checksum How to use checksums |
7bd71f8b0ee0dfbd47ca47f2d514214282806af72aaf71c4761d0432200d086c
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.9.6
|
Security for finee 1.0.3
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.