FinanceGPT - Bilingual Financial Intelligence System
A LangGraph multi-agent system integrating vector DBs, GraphRAG, and Knowledge Graphs for finance data, achieving >95% retrieval accuracy with bilingual EN/UR support.
Client: Financial Services Client (Confidential)
01 Overview
Financial professionals dealing with large document libraries - annual reports, regulatory filings, internal research - spend disproportionate time searching for information they know exists somewhere in their corpus. This client needed a system that could answer complex financial questions across hundreds of documents in both English and Urdu, with citation-level accuracy.
We built FinanceGPT: a multi-agent financial intelligence system that ingests heterogeneous financial data (PDFs, CSVs, structured reports) and enables natural language querying with >95% retrieval accuracy.
02 The Challenge
• Financial data is highly structured but spread across PDFs, CSVs, Excel files, and proprietary formats - each requiring specialized parsing
• Relational financial reasoning (e.g., "compare Q3 revenue across subsidiaries with operating margin > 15%") requires graph-based knowledge representation, not just flat vector search
• Bilingual support (English and Urdu) with correct handling of mixed-language documents and RTL text rendering
• >95% retrieval accuracy requirement - meaning responses must cite the exact source passage, not hallucinate
• Strict data confidentiality - the system runs fully on-premise with no data leaving the client's infrastructure
03 Our Solution
• **Hybrid RAG architecture**: Combines dense vector retrieval (Pinecone) with Knowledge Graph traversal (Neo4j) - complex relational queries route to the graph, semantic search routes to the vector index, and a fusion layer merges results
• **Multi-agent LangGraph system**: A router agent classifies incoming queries and delegates to the appropriate specialist (tabular data agent for CSV/Excel, document agent for PDFs, comparison agent for cross-document analysis)
• **Bilingual embedding model**: Fine-tuned multilingual sentence transformer handling EN/UR code-switching common in Pakistani financial documents
• **Structured citation layer**: Every response is grounded with source document, page number, and passage - citations are verified against the source before delivery
• **On-premise LLM**: Deployed with a locally-hosted Mistral model for data confidentiality, with optional OpenAI fallback for complex reasoning tasks
04 Results & Impact
Project Details
- Category
- Data Science
- Client
- Financial Services Client (Confidential)