Skip to main content

Best Practices for Creating a SignalWire AI Agent

Overview​

When designing an AI Agent with SignalWire, it's crucial to achieve a harmonious balance between clarity, efficiency, and adaptability. The following guide offers a detailed overview of best practices to make sure your SignalWire Agent operates effectively and offers a user-friendly experience.

Crafting the Initial Prompt for the AI​

The initial prompt sets the tone and context for your AI Agent. It's essential to be concise, clear, and relevant.

  • Avoid Over-Prompting: Overloading the AI with excessive information can muddle its understanding. Instead, furnish it with a brief yet comprehensive overview that defines its role and boundaries.
  • Stick to Necessities: Clearly outline the AI's identity and primary responsibilities. This guarantees the AI comprehends its mission and can respond appropriately.
  • Use Markdown for Structuring the Prompt: Structuring prompts in Markdown is highly recommended. It ensures content is organized and legible, both boosting readability and narrowing the AI's interpretation.

Delving into Prompt Configuration​

Understanding the prompt's configuration parameters is key to fine-tuning the AI's responses:

  • Temperature (temperature): Influences the randomness of the AI's output. A value closer to 0 makes the AI's responses more deterministic and focused, while higher values introduce more variety.
  • Top P (top_p): Another parameter influencing randomness. It dictates how diverse the AI's responses can be. A lower value narrows down the potential responses.
  • Confidence (confidence): Determines the threshold for speech-detection events. A lower value reduces the pause after user interaction but may lead to false positives.
  • Presence Penalty (presence_penalty): Dictates the AI's propensity to introduce new topics. A positive value makes the AI more likely to diversify its responses.
  • Frequency Penalty (frequency_penalty): Influences the AI's tendency to repeat itself. Positive values discourage repetition.
sections:
  main:
    - ai:
        prompt:
          text: |
            Your name is David. You are a virtual assistant that is helping with tree planting
            and their environmental benefits.

            ## Information on Tree Planting
            Tree planting is a vital activity that significantly contributes to the environment.
            Trees absorb CO2, prevent soil erosion, provide habitat for wildlife, and offer many
            other benefits.

            ## David's Personality and Job Duties
            Your duties are to help assist users with information related to tree planting.
            You have a friendly and eco-conscious personality.

            ## Greeting Rules
            Greet the user and thank them for showing interest in tree planting. Introduce yourself as David.
            Prefix the greeting with a 'good morning', 'good afternoon', or a 'good evening' depending on the time of day.

          temperature: 0.89
          top_p: 0.64
          confidence: 0.5

Exploring Functions in SignalWire AI​

SignalWire's AI Agent uses functions to neatly store and organize information. This way, the AI can easily handle diverse questions without needing long and complicated prompts.

Employ data_map and webhook_url for Storing Prompts​

Utilizing data_map to store specific prompts or text segments is a practical approach. It allows you to retrieve them when necessary, promoting a fluid and logical flow of dialogue.

Moreover, SignalWire's AI Agent offers the flexibility to incorporate external logic via the webhook_url. By pointing the AI to your own webhook, you can harness backend logic to further customize and refine the AI's responses based on real-time data or specific conditions.

SWAIG:
  functions:
    - function: tree_benefits
      purpose: To share the benefits of planting trees.
      argument:
        type: object
        properties:
          benefit_type:
            type: string
            description: Specific benefit type of tree planting.
      data_map:
        expressions:
          - string: ${lc:args.benefit_type}
            pattern: carbon
            output:
              response: "Trees play a crucial role in absorbing carbon dioxide, thereby helping combat climate change."
    - function: lookup_tree
      web_hook_url: https://example.com/webhook_endpoint
      purpose: To lookup a type of tree.
      argument:
        type: object
        properties:
          tree_type:
            type: string
            description: Type of tree

Optimize Token Usage​

By invoking functions to offer specific guidelines when necessary, you utilize tokens more efficiently, ensuring both optimal AI performance and cost-effectiveness.

Elaboration on the Function's Efficacy​

  • Dynamic Response Generation: The tree_benefits function is designed to provide tailored responses based on user queries. This is evident from the way the data_map is structured. Depending on whether a user asks about the benefits of trees in relation to "carbon", "soil", or "wildlife", the AI will furnish a specific, relevant answer.
  • Pattern Recognition: By using the pattern keyword within the data_map, the function can identify specific keywords or patterns in the user's inquiry. This ensures that the AI's response is not just generic but directly corresponds to the user's query.
  • Token Efficiency: Instead of having a lengthy prompt containing all the possible benefits of trees, the function offers a modular approach. It only provides detailed information about a specific benefit when asked, ensuring that tokens (which represent chunks of information the AI can process in a single go) are used judiciously.
  • Scalability: With the current structure, adding more benefits or details becomes straightforward. For instance, if in the future there's a need to include benefits related to "air quality" or "shade", it can be seamlessly integrated into the function without overhauling the existing setup.
  • Enhanced User Engagement: By furnishing precise, on-demand information, the AI elevates user experience, ensuring interactions are concise and meaningful.

In essence, the tree_benefits function exemplifies the power of modular design in AI prompts. It showcases how, by intelligently structuring functions, one can create an AI Agent that's both responsive and resource-efficient.

Utilizing Perl Compatible Regular Expressions (PCRE)​

The PCRE is a powerful library that provides a set of functions to match patterns using regular expressions. In the context of a SignalWire AI Agent, PCRE can be used to match specific patterns or keywords in a user's response and dictate the AI's subsequent actions.

Benefits of Using PCRE in SignalWire AI​

  • Flexibility: PCRE allows for complex pattern matching, enabling the AI to recognize a wide range of user inputs.
  • Precision: PCRE ensures that the AI's responses or actions are tailored to the user's specific queries or statements.
  • Efficiency: By using regular expressions, developers can write concise patterns that capture a variety of user inputs, reducing the need for long lists of possible phrases or keywords.
  • Scalability: As the system grows or changes, patterns can easily be updated or expanded to accommodate new functionalities without overhauling the entire system.

Example: Using PCRE to Share Benefits of Planting Trees​

Consider the following SignalWire AI configuration that employs PCRE for sharing benefits of planting trees:

SWAIG:
  functions:
    - function: tree_benefits
      purpose: To share the benefits of planting trees.
      argument:
        type: object
        properties:
          benefit_type:
            type: string
            description: Specific benefit type of tree planting.
      data_map:
        expressions:
          - string: ${lc:args.benefit_type}
            pattern: /carbon/i
            output:
              response: "Trees play a crucial role in absorbing carbon dioxide, thereby helping combat climate change."
          - string: ${lc:args.benefit_type}
            pattern: /soil/i
            output:
              response: "Trees help prevent soil erosion and improve soil quality, making the land more fertile."
          - string: ${lc:args.benefit_type}
            pattern: /wildlife/i
            output:
              response: "Trees help wildlife flourish.Many animals also use trees for resting, nesting and for places from which to hunt or capture prey. When the trees mature, animals are able to enjoy delicious fruits and foraging opportunities."
Example Overview:​
  • PCRE patterns like /carbon/i, /soil/i, and /wildlife/i are used to match specific benefit types provided as input.
  • Depending on the matched pattern, the AI provides a relevant response about the benefits of tree planting related to carbon absorption, soil quality, or wildlife support.
  • The /i modifier in the patterns makes the regex match case-insensitive, adding flexibility to user input.

Harnessing the Power of Hints​

Hints serve as beacons, directing the AI's comprehension and making sure it stays relevant to the given context.

  • Keyword Guidance: Introducing specific keywords as hints can channel the AI's attention, ensuring it resonates with the intended theme.
  • Example: For a tree planting and environmental benefits AI, hints like "sapling", "soil", "watering", "benefits", and "environment" are pivotal.
hints:
  - wildlife
  - carbon
  - sapling
  - soil
  - watering
  - benefits
  - environment

Other General Best Practices​

  • Regularly Review and Update: AI, like any other tool, benefits from regular reviews. Periodically assess its performance and make necessary adjustments.
  • Test with Real Users: Before deploying, conduct beta testing with real users to gather feedback and identify areas of improvement.
  • Stay Updated with SignalWire Changes: AI and associated platforms evolve. Stay updated with SignalWire's documentation and update your AI accordingly.
  • Monitor Usage Metrics: Keep an eye on token consumption, user interaction times, and other relevant metrics to optimize performance and manage costs.

Tree Planting AI Example​

This example illustrates a comprehensive AI virtual assistant configuration, named David, designed to guide users on tree planting and its myriad environmental advantages. By harnessing structured YAML configurations, the AI Agent functions effectively, ensuring users receive accurate and pertinent information.

sections:
  main:
    - ai:
        prompt:
          text: |
            Your name is David. You are a virtual assistant that is helping with tree planting
            and their environmental benefits.


            ## Information on Tree Planting
            Tree planting is a vital activity that significantly contributes to the environment.
            Trees absorb CO2, prevent soil erosion, provide habitat for wildlife, and offer many
            other benefits.

            ## David's Personality and Job Duties
            Your duties are to help assist users with information related to tree planting.
            You have a friendly and eco-conscious personality.

            ## Greeting Rules
            Greet the user and thank them for showing interest in tree planting. Introduce yourself as David.
            Prefix the greeting with a 'good morning', 'good afternoon', or a 'good evening' depending on the time of day.
          temperature: 0.89
          top_p: 0.64
          confidence: 0.5
        post_prompt:
          text: '## Post Prompt Actions\n\nSummarize the call and provide the summary in a JSON format.'
          temperature: 0
          top_p: 1.0
        post_prompt_url: https://example.site/webhook_url
        params:
          direction: inbound
          swaig_allow_swml: true
        SWAIG:
          functions:
            - function: tree_benefits
              purpose: To share the benefits of planting trees.
              argument:
                type: object
                properties:
                  benefit_type:
                    type: string
                    description: Specific benefit type of tree planting.
              data_map:
                expressions:
                  - string: ${lc:args.benefit_type}
                    pattern: /carbon/i
                    output:
                      response: "Trees play a crucial role in absorbing carbon dioxide, thereby helping combat climate change."
                  - string: ${lc:args.benefit_type}
                    pattern: /soil/i
                    output:
                      response: "Trees help prevent soil erosion and improve soil quality, making the land more fertile."
                  - string: ${lc:args.benefit_type}
                    pattern: /wildlife/i
                    output:
                      response: "Trees help wildlife flourish.Many animals also use trees for resting, nesting and for places from which to hunt or capture prey. When the trees mature, animals are able to enjoy delicious fruits and foraging opportunities."
            - function: region_specific_trees
              purpose: To recommend trees suitable for specific regions or climate zones.
              argument:
                type: object
                properties:
                  region:
                    type: string
                    description: The region or climate zone for tree recommendations.
              data_map:
                expressions:
                  - string: ${lc:args.region}
                    pattern: /tropical/i
                    output:
                      response: "For tropical regions, consider planting trees like Mango, Coconut, and Banana."
                  - string: ${lc:args.region}
                    pattern: /temperate/i
                    output:
                      response: "For temperate zones, Oak, Maple, and Birch trees are great choices."
            - function: tree_planting_guide
              purpose: To guide users on the tree planting process.
              argument:
                type: object
                properties:
                  step:
                    type: string
                    description: Specific step or aspect of tree planting.
              data_map:
                expressions:
                  - string: ${lc:args.step}
                    pattern: /digging/i
                    output:
                      response: "Dig a hole that is twice as wide as the tree’s root ball and just as deep. Place the tree in the hole and fill it with soil."
                  - string: ${lc:args.step}
                    pattern: /watering/i
                    output:
                      response: "Water the tree immediately after planting. Regularly water it, especially during dry periods."
        hints:
          - sapling
          - soil
          - watering
          - benefits
          - environment

In conclusion, creating an effective SignalWire AI Agent is a blend of clarity, strategic function utilization, and continuous assessment. By adhering to these best practices, you'll ensure that your AI Agent is both efficient and user-friendly.