RAG App

For this project, I engineered a containerized, cloud-native Retrieval-Augmented Generation (RAG) assistant deployed on Hugging Face Spaces that enables users to securely upload documentation and query it in real time through a Streamlit chat interface. Writing native code directly with the modern google-genai and Pinecone Python SDKs allowed me to bypass bulky orchestrator frameworks, resulting in a lightweight, zero-local-storage pipeline that uses text-embedding-004 and gemini-1.5-flash. Navigating this build came with distinct engineering challenges, particularly resolving rigid SDK endpoint routing mismatches and optimizing a vector dimensionality gap by implementing types.EmbedContentConfig to dynamically truncate embedding outputs down to 768 dimensions to match the Pinecone serverless index. I am incredibly proud of how fast, clean, and cost-effective this standalone architecture turned out, as well as the strict system prompt guardrails that prevent hallucinations by defaulting to a clean fallback response whenever a query falls outside the uploaded context. To thoroughly test the live application, users simply need to verify their API keys, paste a technical text block into the ingestion sidebar to index the cloud vectors, and then run a mix of specific domain questions and out-of-domain prompts to observe the system's precise retrieval performance and defensive constraints.