Secure a RAG API without minting tokens or storing passwords
An identity provider issues the tokens, the API only verifies them, and neither the model nor an uploaded PDF gets a say in who the user is.
Series
AI Engineering in Production
Production RAG and agent systems that have to stay up: auth, retrieval, agents with brakes, evaluation and observability. Every post comes from a system I actually built and measured, and I only keep the complicated piece when a simpler one loses on a real metric.
Posts in this series
How to cache an LLM agent with Redis without serving wrong answers
Three caches in front of an agentic RAG system, and five cases where a cache hit must be treated as a miss
Rate limit LLM tokens with Redis when the cost arrives after the call
Reserve an estimate before the model call, correct it after, and give every agent to run its own budget.
Single agent vs multi-agent: check the context before you split
In my RAG eval, [X] of [N] failures never had the evidence in context. A second agent would have failed the same way at about 4x the tokens.
Idempotent Tool Calls: Stop Your AI Agent From Charging Twice
Retries come from four layers and can multiply to 18 runs of one action. Seven rules make a dangerous tool call happen exactly once.
Authorise AI agent tool calls from the JWT, not the prompt
The JWT was valid. The model still picked the customer ID. Move identity into the state so a fooled model can't reach the wrong data.
