Skip to main content
You can copy full or partial flows between projects using Cmd/Ctrl + C and Cmd/Ctrl + V in the Flow Editor.
example-main This flow handles a real-world reservation confirmation scenario. It highlights:
  • Step-specific prompts and function references
  • ASR biasing for input accuracy
  • Clean state transitions between flow steps

​
Start step

To help identify some of the key parts of a flow step, the start step is annotated. Click a lettered heading to go to the relevant section. example-1

​
A. Flow name (breadcrumb)

The flow name appears at the top of the editor and identifies the current flow — in this case, reservation confirmation. It also acts as a breadcrumb, allowing you to return to the list of all flows. This is useful when working across multiple projects or agents.

​
B. Step name and icon

This shows the current step: Collect confirmation code. The 🏁 icon indicates it’s the start step of the flow. Each step contains:
  • A prompt shown to the user
  • Transition logic to guide conversation flow
  • A set of visible functions the LLM can call.

​
C. Transition function reference

Inside the step prompt, the LLM is told to call save_confirmation_code if the user provides a valid input. This example uses both the conversation and flow objects: 
def save_confirmation_code(conv: Conversation, flow: Flow, confirmation: str):
    conv.state.confirmation_code = confirmation
    flow.goto_step("Collect first name")
    return
Transition functions are inserted using / and appear as underlined blocks in the prompt editor. See the transition function page for more details.

​
D. Add another reference

Clicking the . icon below the prompt lets you:
  • Add another function reference
  • Insert a transition to another step
  • Reference rules or knowledge base topics
This is useful for branching logic but avoid chaining multiple goto_step() calls inside a single function. Always return after calling goto_step.

​
E. ASR biasing

This step has ASR biasing set to Alphanumeric, improving recognition of spoken confirmation codes like “B–4–Z–Q–9”. You can enable other biasing types as needed, including:
  • Name spelling
  • Time and date
  • Numbers
  • Addresses
See the ASR biasing page for more details.

​
F. Flow toolbar (bottom panel)

At the bottom of the Flow Editor:
  • Flow functions opens a modal to manage logic used in the flow
  • + Step adds a new node to the flow canvas
Transition functions created here are scoped to the flow unless declared globally.

​
Middle steps

example-2 The agent collects the user’s first and last name using separate steps, each with its own prompt and transition logic.
  • If the user has already provided the name, the agent calls the relevant function (e.g. save_first_name, save_last_name) and proceeds.
  • If not, the agent asks for it directly.
Each step uses few-shot prompting to improve recognition of tricky formats: 
User: It's Smith.
Johannes: Thanks — that's Smith.

User: My surname is de la Cruz.
Johannes: Got it, de la Cruz.

User: Sure, that's H-O-W-E. Howe.
Johannes: Thanks for spelling it — I've got Howe.
ASR biasing is set to Name spelling in the “Collect last name” step to improve recognition of spelled-out inputs.

​
End step

example-3 The agent attempts to match the user’s provided details against active reservations.
  • It compares the confirmation code, first name, and last name against entries in the $reservations list.
  • If a match is found, it calls the confirm_reservation function and moves forward.
  • If no match is found, it calls transfer_call with the following parameters:
    • destination="RESERVATIONS"
    • reason="RESERVATION_NOT_FOUND"
    • utterance="Right, let me put you through to someone who can help. Just a moment."
This is a standard validation step using dynamic values and conditional logic.

​
Reminders

  • LLMs only see the current step’s prompt and function list — not previous steps
  • Always use return after flow.goto_step() to prevent silent overrides
  • Avoid naming transitions by position (goto_step_4) — use intent-based names like match_reservation or retry_lookup

​
Improving date accuracy from speech

Voice transcription (ASR) often produces unstructured date strings like twenty seventh of june twenty twenty five, which can be hard to parse — especially if you’re enforcing a strict format like DD/MM/YYYY.
Instead of checking the date format in the step prompt or rejecting invalid inputs immediately:
  • Move format enforcement into the LLM context field of the parameter.
  • Let the LLM interpret and convert the input to the required format.
  • Then validate it in the function.
This avoids issues where STT fails to cleanly separate day, month, and year — especially in multi-lingual or noisy environments.
Example context (LLM parameter):
The user will say a date. You must extract the intended date and convert it to DD MM YYYY format. If they use a different format, rewrite it before returning. 
try:
    day, month, year = value.split(" ")
except ValueError:
    return {
        "utterance": "Hmm, I didn't catch a full date. Could you say it again?"
    }
This code uses Python’s try statement to attempt splitting the date into day, month, and year. If the input isn’t in the expected format, it prompts the user to try again.