Beacon - An Offline Knowledge Agent

• 

  Sample Conversation 6 - Refusal As Knowledge Outside of Scope

• 

  Sample Conversation 5 - Post Hurricane Clean Up

• 

  Sample Conversation 3 - Safe Water Storage Bihar

• 

  Sample Conversation 1 - Emergency Bleeding Support

• 

  Sample Conversation 4 - Pesticide PPE Bihar

• 

  Sample Conversation 2 - Trains in Bihar

Inspiration

Beacon was inspired by Hurricanes Milton and Helene, which left parts of Florida without internet for weeks, and by Bihar, India, where 53% of people still lack reliable access. When official sites vanish and searches fail, Beacon keeps lifesaving knowledge available through community curated Knowledge Packs.

GPT OSS makes this possible by providing an open, powerful model that runs locally, can call tools reliably, and works fully offline. On its own, GPT OSS cannot know which shelters are open or which local contacts matter. Knowledge Packs close that gap by grounding the model in trusted, community verified sources.

What it does

Beacon delivers trusted answers without internet by drawing only from installed Knowledge Packs. These are local documents, images, and citations curated for a specific region by first responders or NGOs.

• In Pinellas County, Florida it can return evacuation maps, shelter locations, and emergency contacts.
  
• In Bihar, India it can guide someone through first aid problems such as deep cuts, snakebites, or maternal care.
  

Beacon uses a small toolset keep it lightweight. Context retrieval, image retrieval via captions, and knowledge pack metadata are its main tools to give cited answers without hallucination.

How we built it

• Schema: A flexible YAML schema for Knowledge Packs, so they can cover any scenario indcluding disasters, rural villages, and wilderness expeditions.
  
• Storage: FAISS vector stores to embed both text and human-written image captions. Beacon only returns an image if its caption matches the query with high confidence.
  
• Agent: Built with LangChain and LangGraph, limited to three tools — context, getImage, and knowledgeMeta.
  
• Models: Powered by Ollama, running gpt-oss:20b for reasoning and nomic-embed-text:v1.5 for embeddings.
  
• UI: Gradio for a lightweight, offline-friendly chat interface.
  
• Knowledge provenance: Every chunk or image returns its citations, with version, date, and source organization.
  

Challenges we ran into

• Vector stores (FAISS): First time using FAISS; debugging index creation, overwrites, and chunk sizes was tricky.
  
• Schema design: Making the YAML manifest generic enough to handle first aid PDFs, evacuation maps, and image metadata while staying simple.
  
• LangGraph & tool calling: Getting GPT-OSS to reliably choose context and getImage when required took carefully engineered system prompts. We also ran into issues with infinite tool-calling loops, which required guardrails and tighter prompt control to resolve.
  
• UI integration: Streaming tool responses into Gradio, showing images only when relevant, and hiding raw tool calls for polish took iteration.
  

Accomplishments that we're proud of

• Built a working offline AI agent with both text and image retrieval in less than a week.
  
• Created demo-ready Knowledge Packs for Hurricane Response (Florida) and Rural Support (India).
  
• Achieved low hallucination behavior by limiting model outputs to curated packs.
  
• Produced a polished demo UI with map zooms, snakebite guides, and knowledge metadata display.
  

What we learned

• How to build and query FAISS vector stores.
  
• Designing flexible schema manifests that generalize across use cases.
  
• Writing effective system prompts to control tool use in GPT-OSS.
  
• Streaming responses and UI polish in Gradio.
  
• The importance of connecting technical design to human use cases. Shelter, first aid, and wellness were forefront topics in our project.
  

What's next for Beacon - An Offline Knowledge Agent

• Website & distribution: A hub where communities can browse, download, and share Knowledge Packs.
  
• Multilingual support: Integrating Whisper (speech-to-text) and Piper (text-to-speech) for local language deployment in India, Africa, etc.
  
• Community Partnership: Work with EMTs, NGOs, and local governments to author packs.
  
• Testing: Develop standard evaluation scenarios for disasters and rural health, measure accuracy against source docs, and collect feedback from local partners.
  
• Deployment: Distribute Beacon online via website download OR low-cost, solar-powered batteries + laptops in disaster-prone or connectivity-poor regions.
  

Built With

• faiss
• git
• gpt-oss
• gradio
• jupyterlab
• langchain
• langgraph
• nomic-text-embed:v1.5
• ollama
• pypdf
• python-3.10+
• pyyaml
• requests