Aequitas-Fin — AI-Powered Financial Intelligence Agent
A specialised RAG system for analysing complex financial reports — built in iterative milestones. Uses hi-res document partitioning via unstructured.io and a dual-embedding architecture: a local BGE-small lane (384-dim) for low-latency private processing and an OpenAI lane (1536-dim) for high-dimensional accuracy. Milestone 1 benchmarked at 0.816 search score (High Confidence) on CIBC financial data using the local model.
Aequitas-Fin is a RAG (Retrieval-Augmented Generation) system built to analyse complex financial reports — specifically dense PDFs such as CIBC Annual Reports. The system is developed iteratively, with each milestone evaluated against real retrieval benchmarks before the next stage is built.
Architecture
The core innovation is a dual-embedding, named-vector architecture deployed on Qdrant:
- Local lane (local_bge): 384-dimensional vectors using BAAI/bge-small-en-v1.5 via FastEmbed. Designed for low-latency, private processing on Apple Silicon (M3).
- Cloud lane (openai): 1536-dimensional vectors using OpenAI text-embedding-3-small. Designed for high-dimensional industry-standard accuracy.
Documents are parsed using unstructured.io’s hi_res layout strategy, which handles complex financial table structures and multi-column layouts that standard PDF parsers miss.
Milestone 1 Results
Benchmark query: “What was CIBC’s net income for the full year 2025?”
- Local BGE-small: Search score 0.8160 (High Confidence) — correctly ranked contextually relevant financial summaries at the top.
- OpenAI lane: Named Vector structure confirmed ready for real API integration.
Status & Roadmap
- Replace OpenAI placeholder vectors with real API calls
- Develop LLM reasoning layer to synthesise retrieved tables into natural language answers
- Build a front-end dashboard for interactive financial queries
Description
RAG system for financial report analysis using dual-embedding retrieval (BGE + OpenAI) and hi-res document partitioning via unstructured.io.