Configure the chat interface for end users

Decide what happens when the page changes record

By the end of this lesson, you can configure the chat interface for end users and record who ran each turn.

Lesson 6 of 6~25 minScript: 00_setup.sql

Configure display, attribution, and feedback

The region shows tool calls, reasoning and token counts. The audit trail says a background job ran the agent. There is no rating of the answers.

Three attributes fix that.

Show developers more than users

Debug Config is a PL/SQL function that returns JSON. Lesson 3 used it to see error text. It is also an authorization check, because it runs on the server for each request:

if apex_authorization.is_authorized('ADMIN_AUTH') then
  return '{ "show_reasoning": true, "show_tools": true,
             "show_metadata": true, "show_errors": true }';
else
  return '{}';
end if;

An admin sees the tool calls. Everyone else sees the answer and nothing else.

The flags do more than change the display. When a flag is false, the plug-in removes those rows from the response on the server, so the browser never receives them. A user who opens the developer tools finds nothing, because there is nothing there.

Tool names are part of this. SC_LIST_INVOICES is harmless. GET_HR_SALARIES tells a user something about your system that you may not have meant to publish. show_tools is what controls that.

Record who asked the question

Lesson 3 showed the audit trail saying the wrong thing:

ID    STATUS     AUDIENCE  CREATED_BY        APEX_USER  APP  PAGE
5294  completed  db        APEX_PUBLIC_USER

created_by is the database user, audience is db, and there is no APEX user at all. The agent runs in a background job, and a job has no APEX session. The row does not name the person who asked.

Set Session Init Code, in the Agent group:

apex_session.attach(
  p_app_id     => :APP_ID,
  p_page_id    => :APP_PAGE_ID,
  p_session_id => :APP_SESSION
);

That runs inside the job, before the agent. The item values are resolved while the browser request is still alive, so the job knows which session to join.

Send another message and run the same query as lesson 3:

ID    STATUS     AUDIENCE       CREATED_BY        APEX_USER  APP  PAGE
5350  completed  authenticated  ADMIN             ADMIN      777    10
5294  completed  db             APEX_PUBLIC_USER

The run now records the person, the application and the page.

Collect feedback

Set Collect Feedback in the Conversation group.

A thumbs up and a thumbs down appear under the newest answer. There is no dialog, and nothing blocks the page. Only a small share of users will ever press one. It is a quality signal, not a survey.

The chat region on an APEX page with a finished answer, and a thumbs up and thumbs down under it.

One rating is stored for each conversation, on the session:

select session_id, feedback_rating, feedback_comment, feedback_at
  from uc_ai_agent_sessions
 where feedback_rating is not null
 order by feedback_at desc;

uc_ai_agents_api.list_sessions returns the same columns, so “list the conversations with a negative rating” is one query. That query points at the prompts and the tools that need work.

Check it yourself

Send one message as an admin and one as a normal user, then compare what left the server:

select role, count(*)
  from uc_ai_chat_messages
 where session_id = :your_session
 group by role;

With show_tools off, the tool_call and tool_result rows are still in the table — they are never sent to that user. The audit trail stays complete, and the user sees only the answer.

Key takeaways

Debug Config is an authorization function, and the filtering happens on the server: if apex_authorization.is_authorized('ADMIN_AUTH') then ...
Session Init Code with apex_session.attach turns audience = db, created_by = UC_AI into audience = authenticated, created_by = ADMIN.
One feedback rating is stored for each conversation, and list_sessions returns it.