Where answers go wrong
The agent answers from the live state of your systems and from your wording. Both can mislead it. This page lists the failures that can happen and the action for each one.
This page covers the content of a reply. For a symptom in the web UI itself, see Troubleshooting.
How do you route a bad answer?
The answer contains an invented detail
A reply can name a resource, a version, or a number that does not exist. The sentence reads well and fits the context. That is what makes it hard to spot.
Symptom: you cannot find the name or number anywhere in your systems.
Action: open the Tools panel and look for the tool call that produced the detail. If no tool result contains it, the detail is unverified. See Check the agent's work.
The answer covers less than you expected
A reply can cover fewer resources than your question meant. It then reads like a complete answer over an incomplete set.
Symptom: the reply names two of your four clusters, or one of three namespaces.
Action: name the resources yourself in the next message. Then ask "Which resources did you not check, and why?".
The answer uses the wrong time window
A message without a time window leaves the choice to the agent. The chosen window often misses your incident.
Symptom: the numbers do not match the incident you are looking at.
Action: state the start and the end time in your message. Write "between 09:00 and 10:00 today", not "earlier".
The answer describes the wrong resource
Names repeat across clusters and namespaces. The agent can land on the wrong copy of a service.
Symptom: the reply describes a service that behaves differently from the one you watch.
Action: name the cluster and the namespace with the service.
Hyground can also write the name for you. Click the plus icon in the bottom
left corner of the message box. If the menu shows
Subsystems, click it and pick the cluster you mean. Hyground writes the
cluster into your message as @ and its name.
The answer stays generic about your environment
The agent reads background about your environment from the knowledge base. A missing document leaves it with general knowledge only.
Symptom: the reply explains the topic in general terms. It does not name your clusters, your services, or your own procedures.
Action: open the Tools panel and find the knowledge searches. They carry a book icon. If a search returned little, tell your administrator which document is missing. See How the agent uses knowledge.
A tool fails
A tool can fail because a connection to your system is not available. The agent then works without that data. It can still write a confident reply.
Symptom: Show response in the Tools panel shows an error text. Reload the page if a row shows no Show response right after the reply.
Action: send your question again. If the same tool fails again, the connection needs attention. Contact your administrator and quote what the response showed.
A row that has no Show response after a reload is different. The tool ran and found nothing. See Read what the tool returned.
A long session drifts
Every reply uses the whole history of the session. In a long session, old facts compete with your current question.
Symptom: the agent returns to a point you already settled, or mixes two topics.
Action: start a new session. Put the few facts you still need into your first message. See Keep one topic in one session.
When do you contact your administrator?
Contact your administrator in these cases:
- A tool fails the same way after a retry.
- A cluster you expect is missing from Subsystems.
- A document the agent needs is missing from the knowledge base.
Your administrator manages the connections between Hyground and your systems. You cannot repair a connection from the session.
Rate a reply that stays wrong, and name a reason. See Rate a reply.