RAG-Powered Financial Announcement Chatbot

NLP
LLMs
machine learning
finance

A Retrieval-Augmented Generation chatbot for financial announcements with citation-based answers and multi-turn conversation support.”

Author

Manasvi Vardham

Published

March 24, 2026

Project Overview

Financial markets move fast, and critical information often lives in corporate announcements. This project builds a Retrieval-Augmented Generation (RAG) chatbot that allows users to ask financial questions and receive context-grounded answers with citations.

The system is designed for Saudi Stock Exchange (Tadawul) corporate announcements and supports multi-turn conversations, advanced retrieval, and a deployed web interface.

Live App:
Open Financial Chatbot

Key Features

  • Citation-backed answers
  • Multi-turn conversation memory
  • Cross-encoder re-ranking for improved accuracy
  • Persistent vector database
  • Interactive Streamlit UI
  • Fully local execution

How It Works

The chatbot follows a Retrieval-Augmented Generation pipeline:

  1. User submits financial query
  2. Relevant documents retrieved from vector database
  3. Cross-encoder re-ranking improves accuracy
  4. Language model generates response
  5. Sources displayed with citations

This approach reduces hallucination and improves reliability for financial analysis. :contentReferenceoaicite:0

Dataset

  • Domain: Corporate Financial Announcements
  • Documents: 1,800+
  • Content Includes:
    • Dividend announcements
    • Earnings reports
    • Corporate contracts
    • Regulatory disclosures

To improve retrieval quality: - Duplicate rows removed
- Missing summaries removed
- Smart chunking implemented

Technical Architecture

Embedding Model
- sentence-transformers/all-MiniLM-L6-v2

Vector Database
- ChromaDB (Persistent Mode)

Re-Ranking Model
- ms-marco-MiniLM-L-6-v2

Generation Model
- google/flan-t5-base

This two-stage retrieval pipeline significantly improves answer accuracy. :contentReferenceoaicite:1

Conversation Memory

The chatbot supports multi-turn conversations by storing recent interactions.

Example:

User:
“What dividend did Aramco announce?”

Follow-up:
“When was it announced?”

The system maintains context to improve retrieval accuracy.

Deployment

The chatbot is deployed using Streamlit with:

  • Chat-style interface
  • Expandable citations
  • Input validation
  • Public deployment support

Tools Used

  • Python
  • ChromaDB
  • Sentence Transformers
  • Hugging Face Transformers
  • Streamlit
  • Pandas

GitHub Repository

View full project and code:

Code Repository

Impact

This project demonstrates how RAG systems can improve reliability in financial NLP applications by combining retrieval, re-ranking, and generation into a single pipeline.

It highlights practical applications of:

  • Large Language Models
  • Information Retrieval
  • NLP
  • Financial Analytics