Loading...
Loading...
Loading...
Loading...
Loading...
Loading...
Loading...
Loading...
Loading...
From the Clickatell AI Portal, you can configure AI assistants, use our Document Manager to train the assistants on specific information, and you can use Curated Responses to create specific responses rather than relying on AI-generated ones.
Follow these steps to use Clickatell AI:
Launch the Clickatell AI portal (also accessible via My Workspace in the ).
your assistants.
on specific information (upload and sync documents).
to link to (optional),
Create (optional).
Use with Clickatell's and applications.
The Clickatell AI service enables you to build, customize, and train modern AI assistants built on a hosted large language model (LLM) and pre-trained on specific information provided by your business. It lets your customers:
Ask questions and extract accurate, concise, and contextually relevant answers from a chatbot, referencing the information it was trained on.
Interact with a live agent supported by the virtual assistant in Chat Desk.
In addition, the Clickatell AI powers the following features across our products:
Chat Desk
Agent Desk
Chat escalation and alerts
Chat de-escalation
Supervisor Desk
Chat sorting and filtering
Chat Flow
Broadcast Messenger
Assistants are trained on documents that you upload and manage in the Document Manager tab. Each document is linked to an assistant, allowing that assistant to generate accurate and helpful responses based on specific information.
Scenario 1: General assistant uses general documentation for context and calls LLM for the response.
Scenario 2: Assistant uses its own and general documentation for context and calls LLM for the response.
Scenario 3: Assistant uses only its own documentation for context and calls LLM for the response.
Scenario 4: Assistant uses and its own documentation for context. If an answer is found among curated responses, it is sent directly to the end-user, bypassing the LLM.
Check the "Use General Context" tickbox under the advanced settings when to ensure it uses the general assistant's general documentation in addition to its own.
On the Document Manager tab, select the you want to train from the drop-down and click Upload.
Add the files containing the information you want to the selected assistant trained on. You can:
Upload PDF files, plain text files (.txt), and markdown files (.md).
Enter a URL to convert a webpage to a text file (.txt).
You can add multiple files at once:
Once the upload is complete, return to the Document Manager page and click Sync.
After syncing, changes may take up to 5 minutes to reflect in the assistant's answers.
You can add/remove/disable documents at any stage. Remember to sync your changes to ensure the assistant uses the latest information.
You can disable, download, or delete the documents from the Document Manager tab.
For the quickest set up, select Wizard Mode. It provides a guided setup, automatically configuring an assistant with best practice defaults. The Advanced Mode, however, allows for more customization.
Select the assistant type, and give your assistant a name. Provide it with clear behavior instructions.
If you select the Intent Recognition assistant type, you must also select the intent set you want to link to it.
Personalize your assistant.
Upload documents to train your assistant.
You can now either Test or Save your assistant's settings.
Optionally, you can configure advanced settings for your assistant:
Temperature: Controls how "random" or "creative" the generated responses are. A lower value makes the response more focused and predictable. Adjusting the temperature allows users to balance between creativity and coherence, depending on the desired outcome of the text generation.
Top P: A technique that helps generate text that is coherent and diverse, providing a middle ground between deterministic and purely random sampling methods. Lower values (e.g., 0.1) make responses more focused and deterministic. Higher values (e.g., 0.9) allow more creativity and variation. Unlike temperature, Top P adapts to the confidence of the model at each token and filters the output distribution to consider only the most probable tokens (words or subwords).
Amount of chunks to return: Controls how many pieces of relevant information are retrieved in response to your search query. Chunking is essential for handling long documents or conversations, ensuring that the model can process extensive data by breaking it into manageable parts.
Language model memory: If enabled, the assistant will respond using internal memory and not the context provided.
Use general context: If enabled, the assistant will use both its documents and the information available in the general assistant's documents.
The newly created assistant now appears in the Assistants tab and is enabled by default.
The General Assistant is added by default. You cannot disable or delete it, but you can edit certain elements.
On the Assistants tab, select the Edit icon (pencil).
From here you can edit the Name, Type, Instructions, and Advanced Settings for the assistant.
You can also change the status of the assistant (Enabled/Disabled).
Before saving your changes, click the Test button to apply the changes and try it out using the chat widget on the right.
Once you are happy with the changes, click Save to update the assistant accordingly.
On the Assistants tab, select the Test icon.
Use the chat widget on the right to test whether your assistant behaves as you expect.
If you want to change the assistant's setup details, click the Edit button.
An intent recognition assistant is trained to identify specific intents from user inputs and uses a reasoning process to ask questions until all intents are identified.
You must create an intent set first before adding an intent recognition assistant to link it to.
Intent sets are groups of related intents, called intent slots, that must be identified. The system asks questions to gather the information needed to fill the required slots. If it cannot identify all slots in the initial input, it will ask follow-up questions until it reaches the predefined loop limit.
To create a new intent set:
Click New Intent Set on the Intent Recognition tab.
Add a name and description for the intent set. These are for internal reference only so you can easily identify them.
Add intent slots. These are specific pieces of information that must be identified to fulfill an intent.
Name: Enter a name for the intent slot.
Set a loop limit (max. 20) to prevent infinite loops. This limit defines the maximum number of attempts the system will make to identify all required slots before displaying the failed attempt message (see below). A reasonable default, 3-5, is recommended to prevent a poor user experience.
Enter a failed attempt message. This message is displayed to the end-user when the loop limit has been reached (see above). This message should guide the user to rephrase, or offer to transfer them to a (human) agent.
Save.
USER: What is the weather in Cape Town tomorrow morning?
The system classifies this as a weather intent.
“Cape Town” is detected as a location slot and "tomorrow morning” as a date_time slot.
USER: “I owe Lisa money”
ASSISTANT: "Do you want to pay her back? How much would you like to pay?"
USER: "Yes, 120."
ASSISTANT: "To confirm, you want to pay 120 dollars to Lisa from your savings account. Is that correct? Tap below to pay now.”
USER: * Taps button *
In this scenario, the bot understood the initial request, and then filled the intent slots with the information gathered by asking a few questions:
the action (pay),
the amount (120),
Assistants are modern chatbots, powered by generative AI, that can be quickly and easily on business-specific information. Assistants can understand customer queries in meaningful ways, perform a wide range of helpful, automated tasks, and answer questions accurately across various domains in a conversational manner.
Multiple assistants can be , each with their own instructions, functions, and settings. Assistants are independent services that can be deployed in Chat Flow, Chat Desk, or even via APIs to other channels. AI assistansts are created in the AI Portal.
Description: The description helps the system understand what to look for.
Examples of expected responses: Provide examples of what responses the system can expect for this slot. This does not have to be a complete list.
Required flag: The required flag indicates whether the slot must be identified for the intent to be considered complete.
the recipient (Lisa)
The assistant gathers this information conversationally and passes it to the back-end, to execute the end-user's request.
A general assistant is added to your account by default.
You can add any number of additional assistants.
A general assistant is added to your account by default. This general assistant can be edited but cannot be disabled or deleted.
The general assistant uses its own set of documentation and curated responses to generate answers. This documentation includes general information about the business, such as its vision and mission, general FAQs, company policies, products and services offered, etc.
You can add as many assistants to your account as you want. Each assistant has their own instruction (system prompt), settings, set of documentation, and curated responses. These assistants can either use their own documentation exclusively or access both their and the general assistants' documentation.
An example includes having different assistants for different departments within a business, such as HR (internal) or Customer Support (external), where each assistant uses documentation specific to its function.
Currently, you can select between two types of additional assistants:
Content Generation
This type of assistant can answer questions or hold general conversations.
Intent Recognition
This type of assistant can classify users' requests, and slot-fill information.
First create an appropriate intent set before creating an intent recognition assistant, as you need to link the assistant to a specific set.
Common or recurring questions from customers can be addressed with hard-coded answers to provide a precise response rather than an AI-generated one when very specific wording is required. The system identifies when a query matches a curated response and provides that as the answer.
In the Curated Responses tab, click New Response.
Select the assistant you want to assign the curated response to.
Enter the question and the exact answer.
Save.
Sync the curated responses to ensure the assistant can access the latest information.
From the Curated Responses tab, you can disable/enable, edit, and delete the curated responses.
Curated responses are also automatically created from . When an agent selects a , the answer is added as a curated response in the Clickatell AI Portal. From here, the response can be edited or deleted, and will only be used after it is synced.
These curated responses are used when a query that matches them come in i.e. if a question comes in that matches a response, the response will be used rather than the system generating a new one. You can review and edit curated responses at any time to improve consistency and accuracy.
This feature uses AI to understand the nature of existing workflows or assistants, and can route the user to the appropriate next step, based on the information available to it.
If a user is interacting with a flow, and changes their mind about what they want to do mid flow, the Context Router is able to route them appropriately.
Configure this feature as follows:
Launch Clickatell AI, and click on Context Router.
Note that the flows and agents listed, are only the ones for which a was generated in Chat Flow. This summary provides the Context Router with the information that it needs.
You can choose to activate or deactivate a flow. If you no longer want users to be routed to a particular flow, you can deactivate the flow (click Sync after), and the router will no longer be active for that flow.
You can test the Context Router. Simply click on Test to simulate user questions and confirm that the user will go to the intended flow.
