Dispatch № 01 · rag-assistant
AI Prototyping Instructor
Full-stack RAG conversational assistant
- Python
- Next.js
- Pinecone
- Azure OAuth
- React
- Vercel
A retrieval-augmented conversational assistant that answers questions strictly from a curated knowledge base, gated behind Azure OAuth and deployed serverless on Vercel.
By the numbers
- Deployed a serverless RAG application on Vercel, architecting a Next.js backend to orchestrate Pinecone vector search and stream responses to a React UI.
- Achieved 94% retrieval accuracy by implementing iterative knowledge-base refinement via query log analysis, with an 80%+ deferral rate on out-of-scope queries.
The problem
Learners needed an on-demand instructor that could answer prototyping questions accurately without hallucinating beyond its source material, and it had to be locked to authenticated organization members.
Architecture
A Next.js backend orchestrates the pipeline: incoming queries are embedded, matched against a Pinecone vector index, and the top-k context is composed into a grounded prompt. Responses stream token-by-token back to a React UI. Access is gated by Azure OAuth so only authenticated users reach the assistant. The whole thing runs serverless on Vercel.
Retrieval quality
I treated retrieval as the product. By mining real query logs I iteratively refined the knowledge base and chunking strategy, reaching 94% retrieval accuracy, and just as important, an 80%+ deferral rate on out-of-scope questions so the assistant says "I don't know" instead of inventing answers.
Stack
Python for the ingestion / embedding pipeline, Pinecone for vector search, Next.js + React for the app, Azure OAuth for auth, Vercel for serverless hosting.