Deploy apps to Vercel

Build offline RAG pipeline with local embeddings and inference

Offline question-answering over confidential documents when zero cloud egress is a requirement.

• · document corpus
• · user query

• · query-relevant chunks with ranks
• · AI response with citations

• · Ollama (local LLM)
• · embedding model (nomic-embed-text)
• · vector database (PGLite)

• · Ollama running locally
• · documents split into chunks
• · embeddings indexed

• · Ollama service down (fail-closed, no egress)
• · poor chunk boundaries degrade retrieval
• · embedding model mismatch between index and query

• · local-only execution (no cloud API keys in prompts)
• · vector search before LLM generation
• · single-writer PGLite for freshness