In the “Settings / AI Alcmeon” section you have the possibility to configure the AI Agent(s) that will generate the texts for your AI.ssistant (suggested responses to the advisor, conversation summary, etc.) or the responses that your AI Bot scenario will send.

For each Agent:

1. Choose an agent type “Use for AI.ssistant” or “Use for AI Bot”
2. Decide whether to use GAIA for this agent, or your own third-party AI engine
3. Add documents to your AI knowledge base: This is the source of everything your AI knows and uses to suggest answers. Link one or more documents to your AI Agent.
4. Set its “prompt”: the tone of its response and the context of your activity

Alcmeon has an open approach to AI , so you can configure AI Agents that use Alcmeon's internal AI engine (GAIA), or plug them into your own AI engine or that of a partner you choose. You can mix these possibilities by having some AI Agents use GAIA and others use your own AI engine.

General Settings

Agent Name

Choose a name for your agent, internal to Alcmeon.

Active

Only active agents can be used for Alcmeon's AI features like AI.ssistant or AI bots.

Use for…

Choose what your agent will be used for: AI agents used for AI.ssistant appear in your AI.ssistant team configuration ; AI agents used for AI Bots appear in the configuration of the "AI Connector" step of your AI Bot scenario .

Please note, once your agent has been created, this choice can no longer be changed.

Agent type

Decide here whether to use GAIA for this agent, or your own third-party AI engine.

GAIA Agent (Generative AI by Alcmeon)

Knowledge documents used

Here, select the knowledge documents that will be used by this AI Agent, if you have already created and enabled any. If not, you can add more later.

To have the documents used (PDF only) added at the end of the answer by your AI.ssistant or your AI Bot as a reference link, you can check the option “Display knowledge documents in the answer”.

Business context

Give your AI Agent some guidance on the context in which it should attempt to formulate a response. This means specifying:

• Who is asking the question : an existing customer? a prospect? another company? …
• On whose behalf your AI Agent is responding : Describe your company and its business in the context of the questions the AI Agent will have to answer

Wizard Context

Provide your AI Agent with details on the role they should take on to respond : a customer service advisor? a salesperson? a legal expert? …

Response style

Detail the format your AI Agent's responses should take. This includes describing the tone of the response, as well as the most important language elements (instructions for starting the response, ending it, taking leave, how to thank, etc.) and the wording or solutions you don't want (refund offer, apologies in certain cases, etc.).

Our recommendation: Keep your response style to no more than 2,000 characters . If you give too many instructions, the AI engine won't be able to follow them all in every case (this is how generative artificial intelligence works).

The rules of an effective style guide

1. Use trial and error

Create and maintain a list of questions that you will use to validate your answer style after each change. This will prevent correcting a minor error in one question from resulting in a major error in another.

2. Write in English

It's best to use English because the other part of the prompt will be in English. If you use parts of another language in the style, those parts are likely to be used in their original language and not translated. Another reason is that since English is the language of instruction for most LLMs, you increase the chances that the instructions will be followed if you use English.

3. Use simple instructions

Try to write simple instructions that pose the least risk of misinterpretation. Don't define too many sub-cases. If you want to add exceptions, place them at the end of the instruction.

4. Order is important

The order of instructions is important. If instructions conflict for a given conversation, there's a good chance the latter instruction will be followed instead of the former. So, it's a matter of finding a balance between ensuring one instruction is followed and the risk that it will be followed instead of another.

Example

Let's take this example of Response Style:

Use the customer name. Thank him for the message, question. If the customer asked for a question, end the answer saying "nous restons à votre entière disposition si besoin". Wish for a good day/night.
Offer to get back "A bientôt à MaMarque".
Add one or tow emojis linked to the theme of the conversation.
If you're not sure of the language, answer in French.

This poses several problems, because in the responses that will be given to the client:

• There will be French like "nous restons à votre entière disposition si besoin" or "A bientôt à MaMarque" even if the conversation with the customer is in another language and for all the rest of the response the AI has translated correctly into the customer's language.
• Emojis will be included in the response even if the customer complains, which isn't necessarily desirable. Therefore, an exception should be added, worded as simply and clearly as possible.
• The rules for formulating the leave will not necessarily be taken into account; they would be better placed in the last instructions.

Corrected response style:

Use the customer name. Thank him for the message or question.
In case the customer asked for a question, end the answer saying we are staying at their disposal for other inquiries.
Add one or two emojis linked to the theme of the discussion, except in the case the user complains about us.
End the message by wishing for a good day / night and offer to come back to the “MaMarque” stores.

Ban toxic questions

This option is enabled by default. Alcmeon AI checks user messages for issues (using moderation models) such as insults, obscene or racist comments . It is possible that this type of message is intentionally sent by a user to generate an unapproved response by Chat Bot and create bad buzz. To avoid this situation, any message deemed toxic is immediately discarded and not passed to the AI.

Anonymize private data

Any data a customer may send in their message is processed by Alcmeon before being sent to an AI engine (such as Mistral, OpenAI, or others). This is very important because there may be GDPR issues if the customer has disclosed some of their personal information in this data.

To anticipate these potential issues, Alcmeon AI detects and replaces private information found in the user's message before sending it to the AI engine, and replaces it with the original information in the AI engine's response before sending it to the client. The following information is detected and replaced:

• Name of the person
• Phone numbers
• Email address
• IBAN code
• Identification code
• Credit Card Number

Other information that could be considered private, such as location (country, city), is actually not replaced, as it is useful in most cases for the AI to correctly answer the customer's question.

When this “Anonymize Private Data” feature is enabled, an additional option becomes visible: “Allow List.” This is a list of terms that you do NOT want replaced during the anonymization process described above. Make sure that you have at least the name of the company, products, and services you offer in this list.

Third Party Agent

A third-party AI agent only requires the configuration of the endpoint URL and API key. This information is mandatory to create such an agent. See all technical documentation on this topic.