Overview

Flows are a powerful tool for guiding deterministic agent behavior. You can use them to gain greater control by breaking down conversations into discrete steps.

​

Overview

A flow is a series of steps that guide the assistant through a structured interaction. Each step in a flow has its own unique prompt, similar to the knowledge base content field. This modular design makes flows particularly useful for scenarios like collecting user information, processing requests for reservations or bookings, or structuring dynamic conversations with complex logic.

Flows are triggered by specific functions or action fields, and once a flow is active, the assistant will progress through its steps sequentially, following defined transition rules.

​

Key features

​

1. Step-by-step prompts

Each step in a flow has a dedicated prompt to keep processing focused. This helps ensure the assistant processes one task at a time, and does not include any superfluous information in its answers. .

​

2. Hidden prompts

Only the current step’s prompt is visible to the model, keeping the conversation on track. This prevents the assistant from “seeing” future steps.

​

3. Transition functions

Transition functions move the assistant from one step to the next. These functions are defined within the flow and can only route to other steps within the same flow.

​

4. Global functions

Functions not tied to specific steps can be used throughout the assistant’s workflows. These functions allow for reusable logic, such as API calls or user input validation.

​

5. ASR biasing

Flows support ASR (Automatic Speech Recognition) biasing to improve transcription accuracy. This is especially useful for collecting alphanumeric data like postcodes or phone numbers.

​

Best practices for flows

​

Keep steps concise

• Ensure each step handles a single cohesive task, like collecting a name or date.
• When handling related pieces of information like the date, time, and number of people for a restaurant booking, combine them into one step for cohesion.

​

Meaningful function names

• Use descriptive names for functions (for example, SaveLastName or CollectPhoneNumber).
• Avoid naming functions based on future tasks (e.g., GoToNextStep), as this can confuse the model.

​

Modular design

• Use flows to encapsulate logic for specific tasks, like booking a reservation or handling a customer query.
• Incorporate fallback mechanisms to handle unexpected inputs.

​

Exiting flows

• Always include an explicit exit function (ExitFlow) to prevent the assistant from looping through unnecessary steps.
• Exiting the flow resets the assistant’s focus, allowing it to handle new user queries effectively.

​

Advanced features

​

Conditional logic

Flows can include conditions to dynamically adapt responses based on user input. For example:

• If a user provides a specific date, the assistant can check availability and guide the user accordingly.

​

Variables in prompts

Flows support the use of variables to dynamically fill in information during interactions. This allows for personalized and context-aware responses.

​

ASR biasing

• Configure ASR biasing for specific steps to improve transcription accuracy for numbers, postcodes, or other alphanumeric inputs.
• Boost keywords to prioritize relevant user inputs during transcription.

​

Example flow

Here’s an example of a flow for making a restaurant reservation:

1. Start Step: Collect the date, time, and number of people.

   • Prompt: “What date and time would you like to book a table for? How many people will be dining?”
   • Function: CollectReservationDetails

2. Check Availability:

   • Call an API to check availability based on user input.
   • If available: Transition to the confirmation step.
   • If unavailable: Prompt the user to choose a different time.

3. Confirmation Step:

   • Prompt: “Great! Your table is booked for [date] at [time] for [number of people]. Is there anything else I can help you with?”
   • Function: ExitFlow

​

Troubleshooting

• Missing exit step: Failing to exit a flow can lead to the assistant repeating unnecessary steps or ignoring new user queries.
• Overloading steps: Including too much information in a single step can overwhelm both the assistant and the user.
• Poorly-named functions: Functions with vague or misleading names can cause unexpected behavior. Remember that when using the PolyAI platform, all title and description fields matter to the LLM.