SupplyGuard

• 

  Supply Guard

I built SupplyGuard to make supplier risk answers immediate and explainable. Too often the data is scattered, so I combined Elastic’s hybrid search (BM25 + kNN) with a conversational layer, so I can ask natural questions and get grounded answers with citations.

I designed a FastAPI backend exposing hybrid search and a RAG-style chat endpoint, backed by Elasticsearch with a dense_vector field for kNN. I created a React UI with a Chat page, risk dashboards, CSV export, and quick “Add Supplier/Certification” flows to demonstrate real interactions. I also wired an optional Vertex AI path that I can toggle via environment variables; when enabled, the chat uses a Gemini provider, otherwise it falls back to a local synthesis so the demo works without cloud creds.

What I learned:

Blending lexical and vector signals improves relevance for short business queries. Grounding conversational responses with retrieved sources increases trust and clarity. On Windows, pinning compatible httpx/httpcore versions and running Uvicorn from the project root prevents import/runtime issues. Hardening Elasticsearch index creation with timeouts makes local demos more reliable.

Challenges I solved:

Python/Windows dependency quirks (httpcore typing issues on 3.14). Index creation timeouts; I added longer client/index timeouts and idempotent creation. Frontend/backend alignment (base URL, trailing slashes) to avoid 307/308 POST redirects.

Team:

Built solo: I designed, implemented, and integrated the entire stack and UX myself.

Built With

• axios
• elastic
• hybrid-search-(bm25-+-knn-over-dense-vector)-frontend:-react-18
• languages:-python-3.11
• lucide-react-(icons)-ai-(optional):-google-cloud-vertex-ai-(gemini)-via-env-toggled-provider-tooling/infra:-docker
• markdown
• pydantic
• tailwind-css
• typescript-backend:-fastapi
• uvicorn-search:-elasticsearch-8.14-(docker)
• vite