Quick Start for Communication Intelligence Engine

Request a number

Before connecting to the Communication Intelligence Engine, you must complete applications for voice qualifications and number resources. Voice number qualification applications, resource requests, and bill viewing are all managed in Voice Service.

1. Apply for voice qualifications: Go to the Voice Service console. In the navigation bar, choose Qualification and Script Management > Qualification Management and click Add New Qualification. Follow the instructions in Enterprise Qualification Management to fill in the required information and submit your application.

2. Apply for voice scripts: After your qualification is approved, in the navigation bar, choose Qualification and Script Management > Scenario & Script Management and click Add Script. Fill in the requested information and script content as prompted. Select Dedicated Mode for the business mode.

3. Apply for number resources: After your script is approved, in the navigation bar, choose Voice Numbers > Real Number Application and click Apply for Number. Request fixed-line and mobile phone numbers for the required regions based on your business needs.

   Note

   The monthly rental fee for a standard number is CNY 35 per month. Charges start in the month the number is activated. Partial months are billed as full months. You can also contact your account manager for fully managed service: Fully managed is a Voice Service feature that automatically applies for and replaces numbers based on your call volume to maintain high answer rates.

Configure an LLM

Create an LLM gateway

1. Go to the Artificial Intelligence Cloud Call Service console. In the navigation bar, choose LLM Communication > Communication Intelligence Engine > LLM Gateway Configuration and click Create LLM Gateway.

2. In the dialog box, enter model details. For example, if using Alibaba Cloud Model Studio:

   Note

   You can use Qwen-series models, open source models, or commercially mature third-party models. If building your own LLM, fine-tune it for your business scenario and configure prompts before connecting to the Communication Intelligence Engine. Your custom LLM must comply with the LLM Gateway Integration Parameter Protocol to connect.

   Configuration item	Description	Best practice setting
   Model name	Enter a custom name to distinguish this LLM gateway.	Test Gateway
   Model endpoint	Enter the LLM invocation endpoint. If using an Alibaba Cloud Model Studio model: https://dashscope.aliyuncs.com/compatible-mode/v1/chat/completions If using an Alibaba Cloud Model Studio application: Replace YOUR_APP_ID in https://dashscope.aliyuncs.com/api/v1/apps/YOUR_APP_ID/completion with your application ID	http://****
   Authorization code	API key for the LLM service. If using Alibaba Cloud Model Studio, go to the Alibaba Cloud Model Studio console to get your API key. If you don’t have one, click Create My API_KEY in the upper-right corner.	s*********************************1
   Model temperature	Sampling temperature that controls the diversity of generated text.	0.7 Note This is a recommended value.
   Model topP	Nucleus sampling probability threshold that controls the diversity of generated text.	0.9 Note This is a recommended value.
   Model topK	Size of the candidate set during generation.	50 Note This is a recommended value.
   Base model version	Model version passed as a parameter to the LLM. For Qwen models, versions include qwen-plus, qwen-turbo, and qwen-max. For more information, see Select a model. Note When using a Model Studio application, do not configure the base model version to avoid test errors.	qwen-plus

Test and publish the LLM gateway

After creating the gateway, test it before publishing for use by LLM applications.

1. Go to the Artificial Intelligence Cloud Call Service console. In the navigation bar, choose Communication Intelligence Engine > LLM Gateway Configuration. Find your gateway and click Test in the Actions column.

2. In the dialog box, enter test content and click Test to view results.

   Results appear in a table with three columns: Base Model Version, Connectivity Test Result, and Test Result, showing the model version used, whether connectivity passed, and the model’s response.

3. After a successful test, click Publish in the Actions column.

Create an LLM application

1. Go to the Artificial Intelligence Cloud Call Service console. In the navigation bar, choose LLM Communication > Communication Intelligence Engine > LLM Application Management and click Create LLM Application.

   Configuration item	Description	Example
   Application name	Enter a custom name.	Test Application
   Model name	Select a published LLM gateway.	-
   Base template	Select a base template.	-
   Maximum concurrency	Default is 10. Contact operations staff to adjust.	-
   Qualification	Select an approved qualification. View qualification details in Voice Service Qualification Management.	-
   Script	Select an approved script. View script details in Voice Service Script Management.	-
   Greeting	Text: Enter custom greeting text. Use variables in the format ${param}. Recording: Upload a greeting audio file in WAV format, up to 10 MB. Note To ensure natural-sounding voice, if you use a recording for the greeting, select a cloned voice matching that recording when choosing voice tone.	Text: Hello, are you ${param}?
   Prompt	Enter custom prompt text. Use variables in the format ${param}. Note A prompt is an instruction given to the LLM to tell it what you want it to do.	You are Xiao Mei, a professional Alibaba Cloud customer service agent handling pre-sales inquiries. Strictly follow these rules: 1. If the customer asks about ${problem}, direct them to view details in ${document}. 2. If the customer asks about ${other}, direct them to view details in ${answer}.
   Call voice tone	Choose a preset or custom voice tone. Adjust voice style, speed, and volume. Enter sample text to preview online.	-

2. Click Submit to finish adding the application.

Configure LLM application properties

1. Select your LLM application and click Conversation Properties in the Actions column.

2. In the dialog box, configure the following parameters:

   Parameter	Description	Recommended configuration
   Silence duration	Specifies the wait time (in seconds) when neither party is speaking before the silence-triggered model is triggered. The range is 3 to 15 seconds.	5
   Silence-triggered model	When the customer is silent, this feature proactively calls the model. The application inputs a fixed phrase to the model: User did not speak. For custom content, contact technical support for personalized configuration. Values: Yes: Enables the silence-triggered model. When neither party is speaking, the model is actively triggered to generate content, and the silence event is pushed to your large model. No: Disables the silence-triggered model.	Yes
   Hang up on silence configuration	Configure the number of silence events that trigger an automatic hang-up. The range is 1 to 5 times.	3
   Smart answering detection	Enabled by default. Automatically detects voice assistants or mailboxes and returns a smart status code via the LlmSmartCallReport-Call Record Message.	-
   Hang up immediately	When smart answering detection identifies a voice assistant or mailbox, you can choose whether to terminate the call. The default is No. You can enable immediate hang-up based on your business needs.	-
   Maximum call duration	The valid range is 300 to 3,600 seconds. The call automatically hangs up when this limit is reached.	-
   Morse code playback configuration	If enabled, an AI-generated Morse code declaration plays at the end of the conversation.	-
   Uninterruptible opening prompt	If enabled, the large model will ignore any interruptions for a specified period during the opening. This is useful for scenarios where an identity or important notice must be fully delivered, preventing the opening prompt from being interrupted by the customer or background noise. Two options are available: Uninterruptible throughout Uninterruptible during a specified opening period	-
   Prevent consecutive interruptions	When the number of interruptions exceeds a threshold, the large model ignores further interruptions for a configured duration and completes its speech to ensure a smooth conversation. You can configure the system to disallow interruptions for X seconds if the number of consecutive interruptions exceeds X.	-
   Acknowledgement phrases	If enabled, when a customer interrupts the large model, it may first respond with natural-sounding acknowledgement phrases like "Go ahead" or "Mm-hmm" to yield the floor to the customer. This reduces the robotic feel and improves the naturalness of the interaction.	-

3. Click OK to finish configuration.

Step 3: Associate Xiaoji Agent configuration (optional)

Prerequisites: You have created a Xiaoji Agent. See Xiaoji Agent.

1. Go to the Artificial Intelligence Cloud Call Service console. In the navigation bar, choose LLM Communication > Communication Intelligence Engine > LLM Application Management.

2. Find your application and click Xiaoji Agent Configuration in the Actions column.

3. Set Associate Xiaoji Agent to Yes and select the corresponding Xiaoji Agent.

   Associating a Xiaoji Agent incurs additional fees. See Billing Details for pricing. If you don’t have a suitable Xiaoji Agent, click Create Xiaoji Agent to create one.

Step 4: Configure message receipts (optional)

1. Go to the Artificial Intelligence Cloud Call Service console. In the navigation bar, choose LLM Communication > Communication Intelligence Engine > Message Receipt Configuration.

2. On the Message Receipt Configuration tab, find your desired message type, enable the corresponding toggle, and complete setup.

   Message queue consumption mode

   1. Enable the toggle next to General Queue.

      After enabling, Default Queue shows the queue name (for example, Alicom-Queue-18...-LlmSmartCallReport). Retrieve message receipts from this default queue. Keep Application-Specific Queue disabled.

   2. After configuration, pull messages using the queue name. For a consumption demo, see Lightweight Message Queue (formerly MNS) Consumption Demo.

   HTTP batch push mode

   1. Enable the toggle next to Message Receiving URL.

   2. Enter your message receipt receiving URL in the input field. The system sends messages to this URL via HTTP requests. Ensure your endpoint returns HTTP status code 200.

      For example, enter https://push.example.com/contextpath/receive.do. On successful test, the response is {"code": 0, "msg": "Success"}. To configure different URLs for different applications, enable the Application-Specific URL toggle.