> For the complete documentation index, see [llms.txt](https://en.help.firstline.cc/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://en.help.firstline.cc/feature/chatbot/scenario/qna-ai-agent.md). # AI Answer Agent ### Overview "AI Reply Agent" is an AI customer service node in the conversation script. It can understand customer questions based on the knowledge content provided by the business, search for relevant information, and generate natural customer service replies. Compared with using fixed keywords or prebuilding every Q\&A pair, the AI Reply Agent can understand different ways of asking and the surrounding context, so it is suitable for: * Frequently Asked Questions (FAQ) * Product inquiries * Delivery and logistics * Membership system * Return and exchange policy * Store and service information * Other customer service questions with clear business knowledge references For example, the knowledge base has already established: "After payment is completed, the item is expected to ship within 3 to 5 business days." When the customer asks: "I paid yesterday. About when will it be shipped?" The AI Reply Agent can understand that the customer is asking about shipping time, search for relevant knowledge, and generate a reply in a natural customer service tone. {% hint style="info" %} The main principle of the AI Reply Agent is to "answer based on the data provided by the business." If the current data is insufficient to confirm an answer, the AI should not treat general knowledge, the customer's statements, or speculation as official business information. {% endhint %} *** ### Possible reply methods The AI Reply Agent will automatically choose the appropriate message format to reply to the customer based on the customer's question, the data found, and the conversation context. | Reply method | Description | | ---------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | **General text reply** | The AI answers customer questions in natural language based on the knowledge content found. | | **Product carousel message** | When the customer asks about a specific product, a product category, or has a product recommendation need, a product carousel can be generated together with the searched products, showing up to **10 products**. | | **Quick reply message** | The AI can provide up to **3 clickable follow-up questions**, allowing the customer to directly choose what they want to learn next. | #### **General text reply** General text is the AI Reply Agent's primary reply method. The AI will generate a reply based on the customer's question and the relevant data found, for example: **Customer:**"How long after paying for your product will it be shipped?" **AI:**"After the order payment is completed, shipment is expected to be arranged within 3 to 5 business days." If a "conversation role" is set, text replies will also be presented with the corresponding tone and style according to the role settings. #### **Product carousel message** When the customer's question is related to products, such as asking about**a specific product, a certain type of product, or wanting to find a product that meets their needs**, the AI Reply Agent can present related products with a product carousel in addition to the text reply. For example, when the customer asks: > "Do you have any products suitable for sensitive skin?" If products matching the question are found, the system can present related products through a product carousel, allowing the customer to view product information directly, showing up to **10 products**. The product carousel can reduce the steps customers need to take to search for products themselves, and extend product inquiries from text replies to product browsing. > The products actually displayed will depend on the product data found in this search. If no matching products are found, a product carousel message may not be generated. #### **Quick reply message** After the AI finishes answering, it can provide up to **3 quick reply options**, guiding the customer to continue asking related questions. For example, after answering a delivery question, you can continue with: * View shipping fee details * Can a delivery time be specified? * How do I check my order status? Customers can directly click an option to continue the conversation without retyping the full question, which helps them explore the next piece of information while reducing typing effort. ### How to use the AI Reply Agent? After entering the "Conversation Script" editing interface, add an **AI Reply Agent**node at the position where AI should handle customer questions. Generally, you can place the AI Reply Agent after events such as "customer sends a message." For example: ``` Customer sends message ↓ AI Reply Agent ↓ Continue with subsequent flow ``` After the customer's message enters the node, the system will look for relevant information according to the node's configured data sources and understand the customer's current question together with recent conversation context. If the knowledge content is sufficient to answer, the system will generate a customer service reply; if no clear answer can be obtained, follow-up processing will be performed according to the node's "no answer" settings. *** ### Feature settings After adding the AI Reply Agent, you can configure functions such as data sources, no-answer handling, conversation role, image analysis, waiting message, and customer feedback.
Setting itemPurposeInstructions
Data sourceDetermine which data the AI can use to find answersYou can choose knowledge base, products, or other supported knowledge sources. The AI will only search within the specified data range for relevant content.
Specify knowledge base categoryLimit the scope of the knowledge base the AI searchesSuitable for businesses with a large amount of knowledge, allowing specific flows to search only related categories and reducing the chance of retrieving irrelevant knowledge.
Default reply when there is no answerSet the reply when the AI cannot confirm an answerWhen no relevant data is found, or the current data is insufficient to form a definite answer, this message can be shown.
Allow AI inference when there is no answerWhen there is no definite answer, still generate a more natural explanation of limitations or next-step guidanceThe AI must still rely on the knowledge that exists currently; it does not mean it can invent nonexistent policies, prices, or services on its own.
Reply with knowledge base share linkAllow AI to reply with URLs from knowledge contentIf the knowledge source contains a complete URL, the AI can provide that URL to the customer as well.
Conversation roleSet the AI customer service role and brand toneControl the tone, wording, and brand style of replies without changing the business knowledge itself.
Enable image analysisAnalyze images uploaded by customersImages will first be analyzed by AI, and the analysis results will be used to help understand the issue and search the business knowledge.
Waiting messageDisplay a prompt while AI is processingYou can set the waiting message content and delayed display time to reduce customers sending repeated messages while waiting.
Enable feedbackCollect customer ratings after the answerOptions such as "Helpful / Not helpful" can be displayed to observe the quality of AI replies.
Feedback messageIntroductory text before the feedback optionsFor example, "Did the above information solve your problem?".
Feedback optionsSet positive and negative feedbackYou can customize the text shown to customers and the feedback value actually recorded by the system.
Use specified variableUse content saved from a previous flow as the AI search querySuitable for advanced flows where data is processed by other nodes first and then the result is provided to the AI Reply Agent.
*** #### 📚 Data sources Set which data sources the AI Reply Agent should search to find information related to customer questions. The following data sources are currently supported: | Data source | Description | Suitable scenarios | | --------------------------------- | -------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------- | | **Knowledge base** | Search articles and content already established in the FIRST LINE knowledge base | FAQ, service policies, operation instructions, delivery and return/exchange rules, etc. | | **Product** | Search established product information | Product descriptions, specifications, product-related questions, etc. | | **Externally uploaded resources** | Search external data that has been uploaded and provided for AI use | Product documents, service instructions, or other supplementary materials | You can choose one or more data sources according to the purpose of the AI Reply Agent. For example, if this node is mainly used to answer**return/exchange, payment methods, and delivery policies**, you can choose "Knowledge base"; if you also need to answer**product information and product specifications**, you can add "Products" as well. **How to choose data sources?** It is recommended to select only the data sources related to the current customer service scenario, so the AI will prioritize searching for answers from relevant content. For example: * **General FAQ**: Choose "Knowledge base" * **Product inquiries**: Choose "Products" + "Knowledge base" * **Product operation issues**: Choose "Knowledge base" + "Externally uploaded resources" If you use the knowledge base, you can also pair it with "Specify knowledge base category" to further narrow the search scope. *** #### 🗂️ Specify knowledge base category If the business knowledge base contains a lot of content, you can further limit the AI to search only specific categories. For example, the knowledge base includes: * Membership system * Product information * Delivery and logistics * Returns and exchanges * Store information * Internal SOP If this conversation script is specifically handling after-sales issues, you can limit the search scope to "Returns and exchanges" and related categories. This can reduce the chance of the AI retrieving content unrelated to the current customer service scenario. For example: | Purpose of the AI Reply Agent | Recommended knowledge scope | | ----------------------------- | ------------------------------------------------------- | | Membership customer service | Membership system | | After-sales service | Returns/exchanges, after-sales service | | Delivery issues | Delivery and logistics | | Product issues | Product information | | General customer service FAQ | Multiple external customer service knowledge categories | {% hint style="info" %} A larger search scope is not always better. When different categories contain similar information but with different applicable conditions, an overly broad search scope may actually increase the risk of incorrect citations. {% endhint %} *** #### ❓ Default reply when there is no answer Set the default message to reply to the customer when AI **cannot obtain a clear answer from the data source**. For example: > "We could not find information that matches your question. You can try asking in a different way or contact customer service for help." It is recommended to provide a clear next step in the default reply, such as asking the customer to restate the question, view other service information, or be directed to a human agent, to avoid the customer dropping off after the AI cannot answer. If "Allow AI inference when there is no answer" is enabled, the system may instead let the AI generate a more natural reply or guidance; if you want the content set here to always be used when there is no clear answer, it is recommended to turn this feature off. *** #### ✨ Allow AI inference when there is no answer Set the default message to reply to the customer when AI **Whether to still use AI to generate a reply when a clear answer cannot be obtained from the data source**. | Settings | Description | Suitable scenarios | | -------- | ----------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------- | | **On** | When there is no clear answer, the AI still generates a reply or provides next-step guidance based on the information currently obtained. | Scenarios such as general customer service, product inquiries, and FAQs where flexible AI replies are acceptable. | | **Off** | When there is no clear answer, do not use AI to generate an inferred reply; instead, display the "Default reply when there is no answer". | Scenarios such as finance, healthcare, legal, and important policies where the reply content needs tighter control. | For example, if the AI finds "For bulk purchases, please contact the business service window" but does not find actual discount information, enabling this feature allows the AI to generate a reply and guide the customer to contact the business service window; when disabled, the configured "Default reply when there is no answer" will be used. **How to choose?** If the customer service scenario allows the AI to provide more flexible explanations or guidance based on existing information, you can enable this feature. If the reply content involves**finance, healthcare, legal, or other high-risk information**, it is recommended to disable this feature. When the data source does not have a clear answer, using the default reply directly or arranging another customer service flow can reduce the risk of incorrect replies caused by speculation or uncertain information. {% hint style="info" %} Enabling this feature does not mean the AI will always obtain the correct answer. If incorrect answers carry a high cost, it is recommended to use more conservative settings and pair them with human customer service or another follow-up process. {% endhint %} *** #### 🔗 Reply with knowledge base share link Set whether the AI should also provide, when answering customer questions, **the public link to the knowledge article generated by the FIRST LINE system**, allowing the customer to open the article and view the full content. | Settings | Description | | -------- | ------------------------------------------------------------------------------------------------------------------------- | | **On** | When the AI answers using a knowledge base article, it can also provide the public share link for that knowledge article. | | **Off** | The AI will not provide the public share link to the knowledge article generated by the system when replying. | For example, a customer asks about the return/exchange process, and the AI uses the "Return and Exchange Policy" knowledge article to generate the answer. After enabling this feature, in addition to directly answering the customer's question, it can also provide the public link to that knowledge article so the customer can view the full explanation. {% hint style="info" %} This setting**only controls the public link to knowledge articles generated by the FIRST LINE system**. If the content of a knowledge article itself contains other URLs, such as an official website, event page, form, or other links manually entered by an administrator, the AI may still provide those URLs in its reply according to the knowledge content, even if "Reply with knowledge base share link" is turned off. Therefore, if there are URLs you do not want to provide to customers, you should directly check and adjust the knowledge article content, rather than relying on this setting alone to restrict them. {% endhint %} *** #### 👤 Conversation role Choose the "conversation role" used when the AI replies. After setting it, the AI will generate replies according to the role's**personality and style settings**, making customer service interactions more lively and presenting a communication style that fits the brand or usage scenario. For example, you can create different roles for different service scenarios: * Friendly and lively brand customer service * Professional and concise product consultant * Relaxed and natural social media editor After selecting a role, the same knowledge content may use different tones, wording, and expression styles to reply to customers depending on the role settings. **When is it suitable to use?** If you want the AI to not only answer questions but also present a clearer brand personality and interaction style, you can set a conversation role. If you value answer consistency and accurate presentation of knowledge content more, it is recommended to avoid setting a role personality that is too complex or too free-form. {% hint style="info" %} Please note:**After using a conversation role, the AI will provide style-based replies according to the role personality, so the generated content may be more expanded than the original knowledge**, and may even deviate from the existing knowledge scope. If used in finance, healthcare, legal, important policies, or other scenarios with high accuracy requirements, it is recommended to reduce the freedom of the role settings and thoroughly test the reply results before formal use. {% endhint %} *** #### 🖼️ Enable image analysis After turning on "Enable image analysis," the AI can recognize the content of images sent by customers and provide relevant replies together with the customer's question. Suitable for: * Product or packaging identification * Product labels and text information * Screenshots of operating screens or error messages * Photos of documents or forms * Other customer service scenarios that need to be understood together with images For example, if a customer sends a product photo and asks, "How do I use this?" the AI can refer to the image content and relevant knowledge to provide an answer. If you want the AI to focus especially on specific information in the image, such as the product name, model, or error code, you can further configure it with an "image analysis prompt." {% hint style="info" %} Please note:**The image recognition results may be affected by image clarity, text size, shooting angle, and content complexity**. If the image information involves important identity, amount, medical, legal, or other content requiring precise interpretation, it is recommended not to rely solely on the AI image analysis result. {% endhint %} *** #### ⏳ Waiting message AI search and answer generation usually takes some processing time. You can set a "waiting message" to show a prompt to customers while the AI has not yet finished processing. For example: "We are looking up relevant information for you, please wait a moment." The waiting message can be configured with: * Message content * Delay seconds The purpose of setting a delay is to avoid showing unnecessary waiting prompts when the AI finishes answering very quickly. If the actual AI response is usually very fast, it is not recommended to set the waiting message too aggressively; otherwise the customer may first receive "looking up" and then immediately receive the answer, increasing unnecessary message volume. *** #### 👍 Enable answer feedback After enabling "Enable feedback," feedback options can be shown after the AI finishes answering, allowing customers to provide opinions on the answer. For example: > Did the above information solve your problem? * Helpful * Not helpful You can combine "feedback message" and "feedback options" to customize the explanatory text and option content shown to customers. The collected feedback can be used later as a reference for reviewing AI reply performance and improving knowledge content. *** #### 📝 Feedback message The "feedback message" is the text displayed before the feedback options. For example: "Did this answer solve your problem?" "Was the above information helpful to you?" It is recommended to use a short, single-purpose question. Not recommended: "Was this answer correct, complete, fast, and in line with your expectations?" Because even if the customer chooses negative feedback, it is hard to determine which part was actually the problem. *** #### ✅ Feedback options You can set the following for positive and negative feedback: * Display text * Actual recorded value For example: | Feedback type | Display text | Purpose | | ------------- | ------------ | --------------------------------------------------------------- | | Positive | Helpful | Indicates that the answer roughly solved the customer's problem | | Negative | Not helpful | Indicates that the answer did not effectively solve the problem | The display text is what customers actually see; the recorded value is used later for system statistics and analysis. *** #### 🧩 Use specified variable By default, the AI Reply Agent uses the customer's current message to search for relevant knowledge. If "Use specified variable" is enabled, you can instead use content already saved by a previous node as the search query. For example: ``` Customer sends message ↓ Preprocessing ↓ Save the processing result as a variable ↓ AI Reply Agent ↓ Search data using specified variable ``` Suitable for the following advanced scenarios: * The customer's input has already been organized earlier * Need to use data obtained from other nodes * Special image recognition flow * Need to pass existing variables in the flow to AI search If there are no special requirements, general customer service Q\&A is recommended to use the customer's latest message directly. *** ### Reply language The AI Reply Agent supports multilingual interaction and will generate corresponding replies according to the language used by the customer, allowing customers in different languages to get help through the same customer service flow. For example, if a customer asks in Traditional Chinese, the AI will reply in Traditional Chinese; if they ask in English, it will reply in English. Therefore, you do not need to build separate AI reply flows for each language; you can use the same AI Reply Agent to handle multilingual customer service scenarios. {% hint style="info" %} The actual reply performance will still be affected by factors such as knowledge content, conversation role, and the customer's wording. If your service audience includes multiple languages, it is recommended to test the reply results in each commonly used language before formal release. {% endhint %} *** ### Notes AI-generated replies may vary depending on the customer's question style, knowledge content, and feature settings. Even similar questions may produce different reply results. To maintain the quality of AI replies, it is recommended to regularly review and update knowledge content, and retest after adjusting the knowledge, conversation role, or AI Reply Agent settings. If applied in**scenarios involving prices, payments, refunds, finance, healthcare, legal matters, or other situations where incorrect information may pose a high risk**, it is recommended to use more conservative settings, avoid AI inference when there is no clear answer, and pair with human service flows as needed. Before formal release, it is recommended to use a "simulation bot" to test real conversation scenarios. In addition to confirming that the AI can answer correctly, you should also test**cases where there is no answer in the knowledge base, follow-up questions, images, multilingual input, and different customer phrasings**to ensure the overall flow behaves as expected. {% hint style="info" %} AI replies may be inaccurate or incomplete. For service scenarios that require high answer accuracy, it is recommended to establish a human verification or transfer mechanism, and not to use AI replies as the sole basis for judgment. {% endhint %}