7 symptoms
Start from what is wrong.
Nobody opens a directory because they want a vector database. They open it because answers are wrong, or slow, or the bill doubled. Pick the symptom; each one is an ordered checklist through the stack, cheapest check first, with the tools worth reaching for at each step and when to skip them.
Band I · Compute · “too slow, too expensive”
Why is my LLM app slow?
A layer-by-layer checklist for LLM latency: where the time usually goes, which check to run first, and which tools help — or do not — at each step.
Why is my LLM bill so high?
Where LLM spend usually concentrates, the order to cut it in, and which tools help at each layer — with the trade-off each cut makes against quality.
Band II · State · “wrong answers”
Why does my LLM give wrong answers?
Diagnose wrong LLM answers by layer: retrieval, prompt or weights. The check to run first, what each fix needs, and why fine-tuning comes last.
Why does my RAG system hallucinate?
Why RAG apps invent facts: thin retrieval, bad chunks, or prompts that allow guessing — and the layer-by-layer order to fix it with the right tools.
Band III · Control · “unreliable, unmeasured”
Why does my AI agent keep failing?
Why AI agents fail in production — timeouts, lost state, unvalidated tool calls — and the layer-by-layer order to make one reliable.
How do I know if my LLM changes actually helped?
How to measure whether a prompt, model or retrieval change improved an LLM app: build an eval set from real failures, score it, and gate changes on it.
How do I stop PII showing up in LLM traces and prompts?
Why customer data ends up in observability tools and prompt stores — and the order to fix it: redact before log, gate at the gateway, validate outputs, then prove it with tests.