---
title: MiniCPM Financial RAG
sdk: gradio
sdk_version: 5.34.0
python_version: "3.11"
app_file: app.py
pinned: false

tags:
  - track:backyard
  - sponsor:openbmb
  - sponsor:modal
  - achievement:offbrand
  - achievement:fieldnotes

---

# 💰 MiniCPM Financial RAG

### 📊 Financial Document Intelligence Powered by Retrieval-Augmented Generation

**Real-world Problem**

Financial reports, insurance documents, annual reports, SEC filings, balance sheets, and investment documents often contain hundreds of pages of complex information.

Finding specific financial insights manually is time-consuming and error-prone.

**MiniCPM Financial RAG** enables users to upload financial PDF documents and ask questions in natural language. The system retrieves the most relevant information from the document and generates accurate, context-aware answers using MiniCPM models.

---

## Build Small Hackathon Submission

Space:
https://huggingface.co/spaces/build-small-hackathon/MiniCPM_Financial_RAG

Demo:
https://youtu.be/0z1i5ESbgYk

Social:
https://x.com/gajanand2004/status/2066422082725163265

Article:
https://huggingface.co/blog/build-small-hackathon/minicpm-financial-rag


---

## 👥 Team

| Name | Hugging Face Username |
|------|----------------------|
| Gajanan Deshmukh | Gaju19 |

---

## RAG architecture

![kids-education-ai](./Architecture.png)


# 🤖 Models Used

| Task | Model | Parameters | Purpose |
|------|--------|------------|---------|
| Financial Question Answering | openbmb/MiniCPM-2B-128K | 2B | Financial reasoning and answer generation |
| Embedding Generation | openbmb/MiniCPM-Embedding-Light | Lightweight | Semantic retrieval and vector search |

---

# 🧠 Why MiniCPM?

| Model | Benefits |
|---------|----------|
| MiniCPM-2B-128K | Lightweight, fast inference, long-context understanding |
| MiniCPM-Embedding-Light | Efficient embeddings with strong retrieval performance |

---

# 🚀 Features

| Feature | Description |
|----------|-------------|
| 📄 PDF Upload | Upload financial reports and PDF documents |
| ✂️ Smart Chunking | Automatically split documents into meaningful chunks |
| 🔎 Semantic Search | Retrieve the most relevant financial information |
| 🧠 Financial Question Answering | Ask questions in natural language |
| 📚 Retrieval-Augmented Generation | Generate context-grounded answers |
| ⚡ GPU Acceleration | Fast inference using Modal GPU infrastructure |
| 📊 Financial Analysis | Analyze revenue, expenses, assets, and liabilities |
| 🎯 High Accuracy Retrieval | FAISS-based vector similarity search |

---

## 🎯 Example Questions

- What is the company's total revenue?
- What is the net income for this period?
- What are the major risk factors?
- Summarize the financial outlook.
- What liabilities are reported?
- What is the operating cash flow?

---

# 📚 Knowledge Pipeline

| Stage | Purpose |
|--------|----------|
| PDF Parsing | Extract text from PDF documents |
| Text Chunking | Break large documents into manageable sections |
| Embedding Generation | Convert text chunks into vector representations |
| FAISS Storage | Store vectors efficiently for retrieval |
| Similarity Search | Retrieve the most relevant document chunks |
| LLM Generation | Generate grounded answers from retrieved context |

---

# ⚙️ Tech Stack

| Layer | Technology |
|--------|------------|
| Frontend | Gradio |
| Backend | Modal |
| LLM | MiniCPM-2B-128K |
| Embeddings | MiniCPM-Embedding-Light |
| Vector Database | FAISS |
| Framework | LangChain |
| PDF Processing | PyPDFLoader |
| Deep Learning | PyTorch |
| Deployment | Hugging Face Spaces |

---

# 📈 Monitoring

The application is continuously monitored to ensure reliability and performance.

| Component | Monitoring Method |
|------------|------------------|
| Hugging Face Space | Build Logs & Runtime Logs |
| Modal Backend | Endpoint Monitoring |
| Retrieval Pipeline | Context Validation |
| Vector Search | Similarity Search Accuracy |
| Question Answering | Response Validation |
| System Health | Runtime Monitoring |

---

# 🌐 Deployment Architecture

```text
Hugging Face Spaces
        │
        ▼
Gradio Frontend
        │
        ▼
Modal Backend
        │
 ┌──────┴──────┐
 ▼             ▼

MiniCPM QA   FAISS Retrieval
```

---

# 📂 Project Structure

```text
MiniCPM_Financial_RAG/

├── app.py
│   
├── requirements.txt
├── Architecture.png
└── README.md
```

---

# 💻 Installation

## Create Virtual Environment

```bash
python -m venv venv
```

## Activate Environment

### Windows

```bash
venv\Scripts\activate
```

### Linux / Mac

```bash
source venv/bin/activate
```

## Install Dependencies

```bash
pip install -r requirements.txt
```

---

# 🚀 Run Application

```bash
python app.py
```

Application URL:

```text
http://localhost:8000
```

---

# 🎯 Target Users

| User Type | Use Case |
|------------|----------|
| Financial Analysts | Analyze reports and statements |
| Investors | Extract investment insights |
| Accountants | Review financial data quickly |
| Auditors | Validate financial information |
| Researchers | Analyze large financial documents |
| Students | Learn financial concepts interactively |

---

# 📊 Benefits

| Benefit | Description |
|----------|-------------|
| Faster Analysis | Reduce manual document review time |
| Accurate Retrieval | Retrieve the most relevant financial information |
| Context-Aware Answers | Grounded responses from document content |
| Scalable Architecture | Handles large financial reports efficiently |
| Cost Effective | Uses lightweight MiniCPM models |

---