LLM Banking Compliance Assistant

Professional RAG-powered application for intelligent querying of banking and financial regulatory documents.

Overview

Enterprise-grade RAG system combining semantic search, cross-encoder reranking, and GPT-4o-mini for accurate, cited answers to complex regulatory questions.

Key Features:

• Multi-stage RAG pipeline with query reformulation and reranking
• Semantic chunking with BAAI/bge-m3 embeddings (1024-dim)
• Real-time streaming responses with source citations
• PostgreSQL + pgvector for efficient vector search
• Document diversity algorithm ensuring multi-source answers

Tech Stack

Backend: FastAPI, Python 3.12, PostgreSQL + pgvector, OpenAI GPT-4o-mini, BAAI/bge-m3, cross-encoder reranker
Frontend: React 18, TypeScript, Vite, TailwindCSS, shadcn/ui
Infrastructure: Server-Sent Events (SSE), HNSW vector indexing

Quick Start

Prerequisites

• Python 3.12+
• PostgreSQL 15+ with pgvector extension
• Node.js 18+
• OpenAI API key

Installation

# 1. Clone repository
git clone <your-repo>
cd LLM_PRODUCT

# 2. Database setup
createdb llmops_db
psql llmops_db -c "CREATE EXTENSION IF NOT EXISTS vector;"

# 3. Backend setup
cd backend
poetry install
cp env.example .env
# Edit .env with your OPENAI_API_KEY and DATABASE_URL
poetry run python scripts/init_db.py

# 4. Frontend setup
cd ..
npm install

# 5. Launch
./run.sh  # Starts both backend (port 8000) and frontend (port 5173)

Example Queries

Credit Risk Mitigation in Large Exposures

Using CRR Articles 399, 401 and 403, explain how a bank should treat credit risk mitigation techniques when calculating large exposures, and illustrate the process with a step-by-step example. Highlight any supervisory conditions mentioned by the ACPR.

IRRBB Stress-Testing Framework

Design a full IRRBB stress-testing framework consistent with EBA/GL/2022/14 and RTS EBA/RTS/2022/10. Include: (a) the modelling of non-maturity deposit behavioural maturities with the 5-year cap exception, (b) the Supervisory Outlier Test for abnormal sensitivities, and (c) parallel shock scenarios of ±200bp. Identify which assumptions require supervisor approval.

Synthetic Securitisation SRT Assessment

Given Articles 244–245 CRR, assess whether a synthetic securitisation achieves significant credit risk transfer. Include both mechanical tests and qualitative supervisory assessment requirements. Refer to how the Capital Markets Recovery Package modifies SRT conditions for on-balance-sheet STS transactions.

Configuration

Key environment variables (backend/.env):

# OpenAI
OPENAI_API_KEY=sk-...
LLM_MODEL=gpt-4o-mini

# Database
DATABASE_URL=postgresql://user:password@localhost:5432/llmops_db

# RAG Pipeline
TOP_K_RESULTS=8
INITIAL_TOP_K=20
SIMILARITY_THRESHOLD=0.65
RERANK_THRESHOLD=0.3
CHUNK_SIZE=800
CHUNK_OVERLAP=200

# Models
EMBEDDING_MODEL=BAAI/bge-m3
RERANKER_MODEL=BAAI/bge-reranker-v2-m3

API Endpoints

GET  /health                    # Health check
POST /api/chat/stream           # Streaming chat (SSE)
POST /api/documents/upload      # Upload document
GET  /api/documents/            # List documents
DELETE /api/documents/{id}      # Delete document

Project Structure

├── backend/
│   ├── app/
│   │   ├── api/              # API endpoints
│   │   ├── services/         # RAG, embeddings, reranking
│   │   ├── models/           # Database models
│   │   └── core/             # Configuration
│   ├── scripts/              # DB init, document upload
│   └── storage/              # Uploaded documents
├── src/
│   ├── components/           # React UI components
│   └── services/             # API client
└── data/                     # Sample documents

RAG Pipeline

1. Query Reformulation: Expand query with GPT-4o-mini (temp=0.3)
2. Vector Search: Retrieve top 20 chunks via BAAI/bge-m3 embeddings
3. Reranking: Cross-encoder scoring, filter by threshold (≥0.3)
4. Document Diversity: Max 3 chunks per document
5. LLM Generation: GPT-4o-mini with streaming, cited responses