Operations

Admin Panel

Answering Agent

Edit the active prompt

Loading prompt...

Editing Environment

Each environment has its own prompt, so a change can be tried on dev before customers see it.

Promote To

Agent Runtime

Choose models by agent

GCS-backed config

Target Environments

Model Catalog

Current models are environment-specific. New selections apply to the checked environments.

Loaded just now
Agent Current By Environment Set Model Selected For Apply
Answering Agent Customer-facing response model
Guardrails Safety and policy checks
PreRoot Pre-thinking and guidance
PostRoot Post-turn feedback loop
Question Answering Knowledge and content answers
Product Search Product retrieval and ranking
Pipeline Trace Anonymizer PII classification and anonymization

Loading model config...

Agent Runtime

Tune agent runtime values

GCS-backed config

Target Environments

Answering Agent

Response length and how many product cards the answer may show. Current values are environment-specific.

Loaded just now
Response Length Target

Question Answering

Current values are environment-specific. New values apply to the checked environments.

Sources

Loading agent config...

Policy Runtime

Configure guardrails

GCS-backed config

Target Environments

Guardrails

Current values are environment-specific. Add new rules or edit existing ones from the Rule Catalog below.

Loaded just now
Mode

Enforce applies activated rules to the Answering Agent. Observe evaluates and logs matches without changing the answer. Disabled skips guardrail evaluation for the selected environment.

Rule Groups

Unchecking a group disables every rule in that group and shows those rules as unavailable in the catalog below.

Rule Catalog

Showing the rules of Dev. Tick a rule to enable it, or use Edit to change it. Saved rules go live on the next chat message, with no deploy.

0 rules

Loading guardrails config...

Guardrails

Promote changes between environments

Pick where to copy from and to. Every difference in rules, and in which rules or groups are switched on, is listed below. Untick anything you do not want to copy, then promote. Mode and thresholds are not part of this; set those in the Guardrails tab.

New guardrail rule

Add a rule

Describe what the assistant must not do, show a few example messages, and say how it should respond instead. Every field has a short explanation. Use Test the rule on the right before you save.

1 Where should it apply?

The rule is saved to every environment you tick and takes effect there on the next customer message. A good habit: save to Dev, test it in the chat, then save again for Staging and Prod.

2 Name and classify it

Groups bundle related rules so they can be switched on or off together. Type a new name to create a group; it is switched on automatically.

A short label shown under the group in the catalog. It helps people find the rule; it does not change how the rule works.

Risk level Required
Rule ID Assigned automatically when you save

3 What should it catch?

One clear sentence about what the assistant must not do. The Guardrails checker reads it for every customer message to decide whether this rule applies, so be specific.

0 / 600
Examples At least 1

Examples show the checker where the line is. On the left, write a realistic customer message that should trigger the rule. On the right, write a similar message that is fine. Two to five pairs work best.

4 How should the assistant respond?

An instruction to the assistant when the rule applies: what to avoid and what to do instead. Customers never see this text.

0 / 800

The message the customer should get, written in English. The assistant translates it into the customer's language and works it into the answer. It is shown only for High risk rules.

0 / 600
Advanced: how the disclaimer is phrased

What the disclaimer should do, written as an action, for example "decline to advise on X and recommend Y". The assistant follows this instead of repeating the text above word for word, which keeps the whole answer in the customer's language. Leave it empty to pass on the disclaimer text as written.

0 / 400

Search Runtime

Configure brand boosting

GCS-backed config

Target Environments

Brand Boosts

Factors below 1.0 anti-boost, 1.0 is neutral, and factors above 1.0 boost. Saved target IDs come from Product Search.

Loaded just now

Brand catalogs and saved boosts are market-specific.

0 brands

Loading Product Search config...

Safety Runtime

Configure the jailbreak classifier

GCS-backed config

Target Environments

Enable the classifier and set the INJECTION confidence threshold independently for each environment.

Loaded just now

Loading jailbreak classifier config...