Private Financial RAG Platform

The challenge

Analysts had thousands of financial documents (filings, reports, scanned statements, spreadsheets and operational databases) that were effectively unsearchable. Public AI tools were ruled out for confidentiality and compliance reasons, and every answer had to be traced back to its source by hand.

The objective

Give analysts fast, trustworthy answers grounded in the firm's own data across all formats and live SQL databases, with strict tenant isolation, role-based access, full source citations, and the option to run entirely inside the client's environment.

Our approach

We built a retrieval-augmented generation platform with a provider-agnostic AI core. Documents are converted with layout-aware parsing for clean files and vision-LLM OCR for scans, chunked per page, embedded and indexed in a per-organization vector store. An agent routes each question to document RAG, natural-language-to-SQL, tabular search or web search, then streams a source-cited answer, while role-based access ensures users only see context they are entitled to.

How it works

1. 01

   Ingest documents, spreadsheets and live SQL sources through a queue-driven worker.

2. 02

   Convert files with layout-aware parsing or vision-LLM OCR for scans.

3. 03

   Chunk per page, embed with BGE-m3 and index into a per-tenant Qdrant store.

4. 04

   Route queries through a LangGraph agent to RAG, NL-to-SQL, tabular or web search.

5. 05

   Stream a grounded answer with de-duplicated, page-level citations.

The build

Frontend

Nuxt 3, Vue 3, TypeScript, English/Arabic (RTL)

Backend

Django, DRF, Celery, WebSockets

AI

LangChain, LangGraph, BGE-m3, Docling, vision-LLM OCR, hybrid retrieval, on-prem vLLM

Data

Qdrant, PostgreSQL, MySQL/MSSQL, MongoDB, Redis

Infra & security

Docker, Traefik, RabbitMQ, Azure AD SSO, TOTP + WebAuthn

The result