SYNTHETIC DEMO — all logs, policies and cases are invented for illustration. Sportello Accesso is sector practice, not an AccessiBit product or client system.
Sportello AccessoRCA LAB · SYNTHETIC
LLM…
RCA · Multi-agent RAG · Azure AI · No code access

How I do root-cause analysis on a multi-agent RAG chatbot on Azure — without code access.

Work at the behaviour layer: system prompts and turn logs. Every bad answer falls into one of four buckets. The bucket decides who acts — me on the prompt, or the owner on search, index, routes or corpus.

Open the 4 cases Try live Prompt Lab

Usual pattern to ask about: Application Insights → Log Analytics (KQL). Not a claim about this client’s stack.

Four root-cause categories

Every wrong answer falls into one of these four.

A prompt cannot create filters, metadata or documents. Don’t declare retrieval “fixed” when you only stopped an unsupported answer.

Live case lab

Four synthetic Azure-style failure cases

Keys: 1–4 cases · N next step · R reveal

Prompt lab · live edge LLM

A/B replay on frozen synthetic context

Runs on Cloudflare Workers AI over synthetic fixtures — not the client’s Azure, not AccessiBit product data.

Output
Press “Run live replay” to replay.
Call close

Proposed pilot

~10 hours / week at $40/hr — versioned prompts, anonymized cases, before/after eval, written RCA notes. Not the posting’s 30+ hrs and not a silent drop to $20–30.

Three questions for Jacopo:

  1. Who owns the product and who approves prompt changes?
  2. What is the prompt change workflow today?
  3. Can I see one anonymized per-turn trace and its schema?