Refugee Policy assistant Agent

This project is a policy Q&A app for refugee regulations. It combines OpenAi gpt-oss model and retrieval-augmented generation (RAG) to answer questions accurately from uploaded policy documents.

Comment 1

Inspiration

While volunteering at Creshendo, I understood how difficult it is for refugees and asylum seekers to keep up with their rights and the ever-changing policies that affect their lives. Legal frameworks are often complex, scattered across multiple sources, and written in technical language. For people already struggling to meet basic needs like finding shelter, this creates serious barriers such as delayed aid, missed protection opportunities, and increased dependence on overstretched humanitarian workers.
This project was born from the idea that AI can help make rights and policies more accessible, clear, and timely for refugees so as to help them better integrate to our communities.

What it does

The Refugee Policy Assistant Agent makes refugee policies easier to understand and use. It:

  • Lets users ask natural questions about their rights and obligations.
  • Retrieves answers directly from official refugee policies and regulations (e.g., Uganda Refugees Regulations 2010).
  • Provides concise summaries while also linking back to the original legal text for verification.
  • Helps refugees, asylum seekers, NGOs, and volunteers save time and access trustworthy information quickly.

How we built it

  1. Collected official refugee policy documents (Uganda, Kenya, Iran).
  2. Used PyPDF + LangChain to load, split, and clean complex legal texts.
  3. Created vector embeddings with sentence-transformers and stored them in FAISS for fast semantic search.
  4. Built a RAG (Retrieval-Augmented Generation) pipeline, where the OpenAI GPT-OSS 20B model acts as the reasoning engine. it combines retrieved policy text with its powerful natural language understanding to generate accurate, context-aware answers.
  5. Integrated everything into a Streamlit web app (tunneled with ngrok) for an interactive and user-friendly demo.

Challenges we ran into

  • Handling different countries’ frameworks within one system.
  • Limited compute for running larger LLMs.

Accomplishments that we're proud of

  • Designed a working end-to-end assistant that makes policies more accessible.
  • Transformed scattered, difficult-to-read legal texts into user-friendly answers.
  • Created something that can genuinely support NGOs, volunteers, and refugees in their work and daily lives.
  • Learned how to merge AI technology with real-world humanitarian impact.

What we learned

  • How to build and deploy a RAG pipeline using LangChain, FAISS, and Transformers.
  • The importance of source transparency in AI systems dealing with legal content.
  • Practical skills in deploying AI apps with Streamlit tools.

What's next for Refugee Policy assistant Agent

  • Expand to more countries.
  • Add multilingual support so refugees can use the assistant in their native languages.
  • Deploy a stable version for NGOs to test in real-world settings.
  • Build feedback loops with users to improve accuracy and usability.

Built With

  • faiss
  • github
  • langchain
  • ngrok
  • openai-gpt-oss-20b
  • openrouter-api
  • pypdf
  • python
  • sentence-transformers
  • streamlit
Share this project:

Updates