RegBridge: dual-collection RAG — FINMA & private docs

• 

  Team

• 

  Reg_Bridge

• 

  RegBridge: semantic search across private docs and FINMA regulations for Swiss financial SMEs

• 

  Readme

• 

  One Pager ENG

• 

  Memory card of artchitecture

• 

  Demo

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

Inspiration

With 24 years of experience in private banking in Geneva (SG, UBP), I witnessed firsthand how compliance officers spend hours manually searching hundreds of pages of FINMA circulars to answer a single regulatory question. One wrong interpretation can cost a bank millions in fines. I wanted to fix that — and the Qdrant challenge was the perfect opportunity to build it.

What it does

RegBridge is a RAG-powered compliance assistant for Swiss financial SMEs — independent wealth managers, cantonal banks, fintechs, and family offices regulated by FINMA.

• Ask any question in natural language (FR, EN, DE, IT) about FINMA regulations
• Get precise answers always citing the exact source (PDF name + page number) — zero hallucinations
• Upload internal company documents (procedures, policies, contracts)
• Activate cross-search mode to automatically detect gaps between your internal procedures and FINMA requirements
• Admin-protected document library — only authorized users can add or remove internal documents

How we built it

Full stack, 100% free:

• Qdrant Cloud — dual vector collections: finma_compliance (FINMA circulars) and company_documents (internal docs)
• Hugging Face — multilingual embeddings (paraphrase-multilingual-MiniLM-L12-v2, 384 dimensions)
• Groq / LLaMA 3 — ultra-fast LLM generation (llama-3.1-8b-instant, temperature=0)
• LangChain — RetrievalQA pipeline with custom compliance prompt
  • MergerRetriever for cross-collection search
• FastAPI + Ngrok — REST API with 5 endpoints running in Google Colab
• Lovable — professional dark-mode chat interface generated by AI

Data: 3 official FINMA circulars (2016/7, 2023/1, 2011/1) chunked at 500 characters with 50-character overlap, indexed in 4 languages (FR, EN, DE, IT).

Challenges we ran into

• Ngrok browser warning blocking Lovable API calls — solved by adding ngrok-skip-browser-warning header
• Groq model deprecation — llama3-8b-8192 was decommissioned mid-build and replaced with llama-3.1-8b-instant
• LangChain import breaking changes — langchain moved to langchain-classic
• Colab kernel restarts — losing all variables mid-demo; solved by restructuring the server notebook to reload everything from scratch on startup
• MergerRetriever for cross-search — combining two Qdrant collections into a single retriever required careful prompt engineering to distinguish FINMA vs internal sources

Accomplishments that we're proud of

• Built a fully functional RAG agent in under 72 hours with zero budget — no credit card, no paid APIs
• Achieved zero hallucinations — the agent always cites its source or explicitly says "this information is not in the available documents"
• Implemented cross-collection compliance gap detection via Qdrant's MergerRetriever — the first feature of its kind for Swiss banking compliance
• Delivered a production-quality UI via Lovable with admin authentication, document library management, and session history
• Full multilingual support (FR/EN/DE/IT) via per-language PromptTemplate objects — no translation pipeline required

What we learned

• Qdrant's dual-collection architecture is powerful for domain separation — keeping regulatory docs separate from internal docs dramatically improves retrieval precision
• Prompt engineering matters more than model size — a well-crafted compliance persona prompt on LLaMA 3 8B outperformed generic prompts on larger models
• RAG is not just retrieval + generation — the chunking strategy, overlap size, and k-value for retrieval have a massive impact on answer quality
• Building with 100% free tools is entirely viable for production prototypes — the limiting factor is engineering time, not budget

What's next for RegBridge

v2.0 — Q2 2026 : DORA Integration Add a third Qdrant collection for the Digital Operational Resilience Act (EU) — directly relevant to my systems engineering background in hybrid cloud banking infrastructure.

v3.0 — Q3 2026 : GDPR Coverage Full European compliance coverage: FINMA + DORA + GDPR in a single platform with a regulation selector in the UI.

Production roadmap:

• Replace Ngrok + Colab with proper cloud deployment (Azure, aligned with hybrid on-premise/cloud banking infrastructure)
• Add Supabase Auth for proper multi-user admin management
• Implement compliance audit trail — log every question, answer, and source for regulatory reporting
• Extend multilingual support for all Swiss cantonal institutions

RegBridge started as a hackathon prototype. It is now becoming the foundation of a compliance product for European financial SMEs.

Built With

• 3
• api
• colab
• database
• face
• fastapi
• google
• groq
• hugging
• langchain
• language
• llama
• lovable
• natural
• ngrok
• pypdf
• python
• qdrant
• rag
• rest
• sentence
• transformers
• uvicorn
• vector