Skip to main content

How to configurate your agent

Step 1 - Set security

PII Detection​

is the process of identifying personal data—such as names, addresses, IDs, or financial details—that can uniquely identify an individual. It helps protect privacy and prevents sensitive information from being exposed. Organizations use PII detection to comply with data protection laws and secure their systems.

Within the platform, choose to enable or disable the fields you want to block

  • Email Addresses
  • Phone Numbers
  • Names
  • Physical Addresses
  • Social Security Numbers
  • Credit Card Numbers
  • IP Addresses
  • Passport Numbers
  • Driver's License Numbers

Prompt Injection Detection​

Detect and prevent prompt injection attacks to maintain AI safety.

Scope Guardrails​

Scope guardrails help you keep the agent focused on what it should handle, and block or redirect questions that are out of scope.

You define the scope in plain language, and the platform uses it to decide when a request is relevant or not.

Scope Definition​

Scope Definition is a free‑text description of what this agent is allowed (and not allowed) to do.
You can write it in any format (sentences, bullets, examples).

We recommend including:

  • What this agent should handle
    • Example: “Answer questions about our billing plans, invoices, refunds, and payment methods.”
  • What this agent should not handle
    • Example: “Do not answer product technical questions, HR questions, or legal questions.”
  • Target audience or context
    • Example: “This agent is only for existing paying customers.”

How it is used​

  • For in‑scope questions, the agent responds normally.
  • For out‑of‑scope questions, the platform can:
    • Block or redirect the request, and/or
    • Instruct the agent to answer with a safe “out of scope” message (depending on your security settings).

This helps prevent the agent from “doing too much” or answering questions it shouldn’t handle.

Step 2 - Set translation

Translation lets this agent talk naturally with users in their own language, while still thinking and working in English behind the scenes.

When enabled, the platform can:

  • Translate the user’s message into English before sending it to the agent.
  • Translate the agent’s answer back to the user’s language before showing it.

Domain​

Domain is the main topic or context of your content (for example: sports, finance, healthcare).

Why it matters:

  • It helps the translation system choose better words and phrases.
  • It reduces mistakes on domain‑specific terms (like sports jargon, medical terms, etc.).

Example:

  • sports – Better handling of team names, positions, stats, and sports terminology.
  • support – Better handling of tickets, accounts, subscriptions, etc.

Choose the domain that best matches what this agent usually talks about.

Source Language​

Source Language is the main language your users write in when talking to this agent.

  • Example:
    • If most users write in Hebrew, set Source Language to Hebrew.
    • We will translate:
      • Hebrew → English before sending to the agent.
      • English → Hebrew when returning the answer to the user.

If your users speak more than one language, pick the primary one for this agent, or create separate agents per language if needed.

Step 3 - Set inbound conversation

These settings control how user messages are translated into English before they are sent to the agent.

Model​

Example: openai/gpt-4o-mini, google/gemini-pro

  • What it is: The AI model we use specifically for translation of incoming messages.
  • How to choose:
    • Use the recommended default (for example: openai/gpt-4o-mini) unless your team tells you otherwise.
    • If your company has a preferred provider (OpenAI, Google, etc.), pick the matching model name.

This model is only used to translate the user’s message, not to generate the final answer (unless your setup says otherwise).

Temperature​

  • What it is: A number that controls how “creative” or “strict” the translation is.
  • Range: Typically from 0 to 1.
    • 0 = very strict and literal (best for accurate translation).
    • Higher values (e.g. 0.7) = more variation and looser wording.

For translation, we recommend keeping Temperature = 0 so the meaning stays as close as possible to the original text.

Max Tokens​

  • What it is: The maximum length of the translated text, measured in “tokens” (pieces of text).
  • Simple way to think about it: A safety limit on how long the translated message can be.
  • Default example: 1000 tokens is usually enough for normal user messages and short paragraphs.

If your users regularly send very long messages or documents, you can increase this value carefully.

Translation Examples​

You can provide example pairs of how certain phrases should be translated, for example:

  • Hebrew → English
    • “מנוי חודשי” → “monthly subscription”
    • “כרטיס מנוי” → “membership card”

Why this is useful:

  • It teaches the translation system how to handle your specific terminology.
  • It helps keep important product or domain terms consistent across all translations.

To add examples:

  1. Click “Add Example”.
  2. Enter the source phrase (e.g. in Hebrew).
  3. Enter the desired English translation.
  4. Repeat for any key terms or phrases that matter to your product.

Over time, these examples help the system better match your brand language and domain.

Step 4 - Enhancement

Enhancement settings help the agent better understand your domain language and react intelligently to specific situations in the conversation.

Terminology​

Use Terminology to teach the AI your domain‑specific terms and how to treat them.

What it does​

Many industries use special words, slang, or abbreviations (for example: “hattrick”, “GOAT”, internal product names).
Terminology lets you:

  • Mark important terms so the AI recognizes them as special concepts.
  • Keep translations and answers more accurate around those terms.

How to configure​

  1. Under Terminology, click “Add Term”.
  2. Enter each important term, one per line.
    • Example terms:
      • hattrick
      • goat
      • Your product or feature names
  3. Optionally (if supported), you can also add a short explanation for internal clarity, like:
    • hattrick – when a player scores three goals in a match
    • goat – greatest of all time (sports slang)

These terms help the AI avoid mistranslations or misunderstandings for key domain words.

Terminology Similarity Threshold​

  • What it is: A number that controls how “strict” we are when matching terms.
  • 0 = exact match only
    • The AI only treats the word as a term if it matches exactly (same spelling).
  • Higher values (e.g. 1, 2) allow small spelling differences or typos.
    • Useful if users often misspell terms.

If you are not sure, you can keep the default value.
For very sensitive terms (like legal/product names), you may prefer 0 (exact match).

Conditional Instructions​

Use Conditional Instructions to set up “if this, then do that” rules, written in natural language.

What it does​

You describe a condition in plain language, and when that condition is detected in the user’s message or conversation, the system can apply special instructions to the agent.

Examples of conditions:

  • “the user is speaking badly about the Barcelona football team”
  • “the user seems frustrated or angry”
  • “the user is asking about canceling their subscription”
  • “bananas are mentioned”

You can then attach special behavior, such as:

  • Answer more calmly and empathetically,
  • Escalate to a human,
  • Avoid certain topics,
  • Use a specific style of response.

(The exact behavior depends on how your team configured instructions for each condition.)

LLM Model​

Example: openai/gpt-4o-mini, google/gemini-pro

  • What it is: The AI model used to evaluate the condition (not to answer the user).
  • It reads the conversation and decides whether your condition is true or false (for example: “Is the user frustrated?”).

You can usually keep the default model unless your team has a specific preference.

LLM Temperature​

  • What it is: Controls how consistent the condition detection is.
  • 0 = deterministic (most stable, less randomness).
  • Higher values (e.g. 0.5) make it more flexible but may be less consistent.

For condition checks (yes/no logic), we recommend keeping Temperature = 0 for predictable behavior.

How to add a condition​

  1. Click “Add Condition”.
  2. Describe the condition in natural language, as if you’re explaining it to a colleague.
    Example:
    • the user is speaking bads on Barcelona football team
    • the user seems frustrated about a recent loss
  3. (If your UI supports it) Attach or configure what instructions should be applied when this condition is true.

The system will then have the LLM model check each message against your conditions and adjust behavior accordingly.

Step 5 - Generation

These settings control how the agent actually writes its answers.

Model​

Example: openai/gpt-4o, openai/gpt-4o-mini, google/gemini-pro

  • The AI model that generates the final response to the user.
  • Keep the default unless your team instructs otherwise or you need a cheaper/faster model.

Temperature​

  • Controls how “creative” vs. “precise” the answer is.
  • 0 = very deterministic/strict; 1 = very creative/varied.
  • For support/accuracy-focused agents, use low values (e.g., 0 – 0.3).
  • For more brainstorming/creative tone, increase slightly (e.g., 0.5 – 0.7).

Max Tokens​

  • The maximum length of the model’s answer (in tokens ≈ word pieces).
  • Bigger number = longer potential answers, but also higher cost/latency.
  • Example: 4000 is usually enough for detailed responses; reduce if you want shorter answers.

System Prompt​

  • The core “instruction sheet” for the agent.

  • Write in plain language what the agent should and should not do:

    Who the agent is (role/tone).

    What topics it should focus on.

    What it must avoid (compliance, brand voice, forbidden topics).

    How to answer (style, length, language).

  • Keep it concise and specific; add bullets for clarity.

Step 6 - Validation

Validation makes sure the agent’s answers follow your rules before they are shown to the user.
You describe what you want to check in natural language, and an AI model reviews the response.

Model​

Example: openai/gpt-4o-mini, google/gemini-pro

  • The AI model used only for checking the agent’s answer against your rules.
  • It does not generate the main answer – it only evaluates it.
  • You can usually keep the default (e.g. openai/gpt-4o-mini).

Max Tokens​

  • The maximum length of text the validation model can use for its own internal reasoning/answer.
  • Default like 1000 is usually enough for normal‑length responses.

Temperature​

  • Controls how consistent the validation result is.
  • For validation, we recommend 0 so the behavior is as predictable as possible.

Validation Rules​

Each rule is a simple question you ask about the agent’s answer, plus what should happen if the answer breaks that rule.

How a rule works​

For each rule, you define:

  • Question
    A yes/no question in natural language that the validation model answers about the response.
    Example:
    did the agent response with negative ideas against Barcelona?

  • Expected Answer
    What you consider “OK” for this rule.

    • If you expect the answer to be “No”, then:
      • If the model answers “No” → the response passes this rule.
      • If the model answers “Yes” → the response fails this rule.

  • Mode
    What to do when the rule fails:

    • Blocking – Block the original answer and show a fallback message instead.
    • Disclaimer – Keep the answer, but add a disclaimer text to it.

Example: Sentiment about a team​

  • Question:
    did the agent response with negative ideas against Barcelona?
  • Expected Answer:
    No
  • Mode:
    Blocking or Disclaimer (see below)

If the validation model decides the answer is negative about Barcelona:

  • The rule fails.
  • The platform either:
    • Blocks the answer and shows a fallback message (Blocking), or
    • Shows the answer but adds a disclaimer (Disclaimer).

Disclaimer​

Disclaimer Text is added to the agent’s response when:

  • A rule fails, and
  • The rule’s mode is set to Disclaimer.

Example disclaimer:

“sorry that I spoke badly about the team - I really admire it.”

This text will be appended to the response so users understand the system is correcting itself.

Use Disclaimer mode when:

  • You still want to show the original answer,
  • But you need to clearly add context, apology, or clarification.

Fallback Message​

Fallback Message is shown when a rule with Blocking mode fails.

  • Instead of the original agent answer, the user sees this safe message.
  • It should be generic, polite, and not expose internal details.

Example fallback message:

“Sorry, this answer did not meet our guidelines. Please try rephrasing your question or contact support.”

Use Blocking mode and a fallback message for cases where:

  • The answer is not safe,
  • Or it strongly violates your policy (e.g. hate, abuse, disallowed topics).

Step 7 - Outbound Translation

Outbound translation controls how the agent’s English answer is translated back to the user’s language before it is shown.

For example:

  • The agent thinks and answers in English.
  • Outbound translation converts that answer into Hebrew for the user.

Model​

Example: openai/gpt-4o-mini, google/gemini-pro

  • The AI model used specifically to translate the agent’s response from English to the target language.
  • You can usually keep the default (e.g. openai/gpt-4o-mini) unless your team prefers another provider.

Temperature​

  • Controls how “creative” or “strict” the translation is.
  • For translation, we recommend 0 so the meaning stays as close as possible to the original English answer.
  • Higher values (e.g. 0.5) would allow looser, more free‑style translations, but are rarely needed.

Max Tokens​

  • The maximum length of the translated answer, measured in tokens (pieces of text).
  • Simple way to think about it: a safety limit on how long the translated response can be.
  • A value like 1000 is usually enough for typical agent replies.

If your agent sometimes returns long paragraphs or summaries, you can increase this limit.

Translation Examples​

Use Translation Examples to teach the system how to translate important phrases or domain terms from English to your target language (for example: English → Hebrew).

This helps:

  • Keep key terms consistent (product names, roles, sports terms, etc.).
  • Avoid awkward or incorrect translations for words that matter to your brand.

To add examples:

  1. Click “Add Example”.
  2. In the English field, enter the original phrase.
  3. In the Hebrew (or target language) field, enter the exact translation you prefer.
  4. Repeat for your most important phrases and terms.

Over time, these examples help the translation system better match your tone, terminology, and domain.