FinSight RAG

FinSight RAG is a Retrieval-Augmented Generation (RAG) application combined with financial sentiment analysis, designed to help users analyze and query financial reports of companies or entities. The system leverages Large Language Models (LLMs), AI agents, and domain-specific NLP models to deliver context-aware, sentiment-driven insights grounded directly in source documents.

---

🚀 Project Overview

FinSight RAG enables users to:

• Upload and store financial reports (e.g., annual reports, earnings statements)
• Ask natural language questions about the content
• Retrieve relevant document chunks using vector search
• Generate grounded answers using LLMs
• Perform financial sentiment analysis on extracted statements
• Interact with the system via a Gradio-based UI

The project combines RAG pipelines, agent-based reasoning, and fine-tuned financial NLP models for robust financial analysis.

---

🔨 Main Tools

• LangGraph – Agent orchestration
• LangChain – Retrieval-Augmented Generation (RAG)
• Hugging Face Transformers – Financial sentiment classification
• Gradio – Web-based user interface

🧠 Core Features

• Retrieval-Augmented Generation (RAG)

  • Semantic document retrieval using embeddings
  • Context-aware answer generation

• Financial Sentiment Analysis

  • Fine-tuned DistilBERT model on financial text

• Agent-Based Reasoning

  • Multi-hop RAG using LangGraph agents

• Interactive UI

  • Gradio-powered web interface

---

🏗️ Project Structure

root/
├── .venv/                # Virtual environment
├── .vscode/              # VS Code configuration
├── data/                 # Documents and datasets
├── models/               # Saved and fine-tuned models
├── finsight_rag/         # Core application package
│   ├── agent/            # LangGraph agent logic
│   ├── config/           # Configuration files
│   ├── datasets/         # Dataset download scripts
│   ├── ingest/           # Document ingestion
│   ├── llms/             # Functions returning LLM instances
│   ├── rag/              # RAG pipelines and retrievers
│   ├── sentiment/        # Sentiment analysis module
│   ├── train/            # Model training scripts
│   ├── app.py            # Application entry point
│   ├── utils.py          # Shared utilities
│   └── tests/            # Unit and integration tests
├── pyproject.toml        # Project dependencies
├── poetry.lock           # Locked dependency versions
└── .gitignore

---

🧪 Models & Technologies

🔍 Embeddings

• sentence-transformers/all-MiniLM-L6-v2

Used for:

• Semantic document chunking
• Vector-based retrieval

---

📈 Sentiment Analysis Model

• distilbert-base-uncased
• Fine-tuned on Financial PhraseBank

Purpose:

• Classify sentiment in financial statements
• Capture domain-specific financial tone (positive, neutral, negative)

---

💬 Language Models (Chat Models)

• meta-llama/Llama-3.1-8B-Instruct
• gemini-2.5-flash

Used for:

• Answer generation
• Multi-step reasoning
• Agent decision-making

---

🧠 AI Agents

• LangGraph

  • Orchestrates reasoning steps
  • Coordinates RAG retrieval + sentiment analysis
  • Enables structured, explainable workflows

---

🔗 RAG Framework

• LangChain

  • Document loaders
  • Chunking & indexing
  • Retrieval pipelines

---

🖥️ User Interface

• Gradio

  • Simple and interactive UI
  • Chat-style financial analysis interface

---

⚙️ Installation

git clone https://github.com/your-username/finsight-rag.git
cd finsight-rag
poetry config virtualenvs.in-project true
poetry install
poetry env activate

---

▶️ Running the Application

python finsight_rag/app.py

Access the UI at http://localhost:7860.

---

🧪 Training the Sentiment Model

python finsight_rag/train/train_sentiment.py

---

📚 References

• Financial PhraseBank Dataset
• Hugging Face Transformers
• LangChain
• LangGraph
• Sentence Transformers

---

📄 License

This project is licensed under the MIT License.

---