Multi_Source_Financial_RAG

Multi-Source Financial Intelligence RAG System

An AI-powered financial research assistant that retrieves relevant evidence from uploaded financial documents and generates citation-backed answers using Retrieval-Augmented Generation (RAG).

---

📌 Overview

Financial documents such as filings and earnings transcripts can contain hundreds of pages of information, making it difficult to quickly locate relevant evidence.

Multi-Source Financial Intelligence RAG System addresses this problem by combining document processing, semantic search, vector retrieval, and large language models into a single research workflow.

Users can:

• Upload financial documents
• Extract and process document content
• Search for relevant evidence using semantic retrieval
• Ask natural-language financial questions
• Receive answers grounded in retrieved document content
• Trace answers back to their source documents and chunks

The system uses a local file-backed architecture with Qdrant for vector search and Gemini for embeddings and answer generation.

---

✨ Features

📄 Document Processing

• Upload PDF, TXT, and Markdown documents
• PDF extraction using PyMuPDF
• pdfplumber fallback for PDF extraction
• Text normalization and cleaning
• Basic financial metadata extraction
• Source and section metadata preservation

🧩 Intelligent Chunking

• Structure-aware document chunking
• Overlapping chunks
• Paragraph and sentence boundaries
• Configurable token limits
• Character-range provenance tracking

🔎 Semantic Retrieval

• Gemini-generated embeddings
• Local embedding cache
• Normalized vector representations
• Qdrant vector search
• Metadata-based retrieval filters

🤖 RAG Question Answering

• Retrieves relevant document evidence before generation
• Gemini-powered answer generation
• Context-grounded responses
• Source citation metadata
• Conversation history support

🖥️ Web Application

• React + TypeScript frontend
• Vite development environment
• Dashboard interface
• Document upload and library
• Semantic search interface
• Research chat
• Dark-mode support

Note: Dashboard summary cards are currently presentation placeholders. Upload, document, search, and chat workflows are connected to the backend APIs.

---

🏗️ Architecture

                    React + Vite Frontend
                              │
                              │ HTTP / JSON
                              ▼
                         FastAPI API
                              │
             ┌────────────────┼────────────────┐
             │                │                │
             ▼                ▼                ▼
       Document            Retrieval          RAG
       Ingestion            Pipeline         Pipeline
             │                │                │
             ▼                ▼                ▼
         Chunking          Qdrant           Gemini
             │            Vector Search       LLM
             ▼                │                │
        Gemini             Retrieved          │
       Embeddings            Context          │
             │                └────────┬───────┘
             ▼                         ▼
       Local Storage            Answer + Citations

---

👤 Author

Aryan Palla

• GitHub: @aryanpalla198