3.2 — Configure LLM, Python, and Prompt Nodes

LLM Node — Required Configuration

When you add an LLM node you must configure:

1. Connection — an Azure OpenAI connection from your project's connected resources (e.g., Default_AzureOpenAI)
2. API type — chat (chat completions endpoint, preferred for GPT-3.5-Turbo, GPT-4) or completion (legacy text completions)
3. Deployment name — the deployment identifier you set when you deployed the model
4. Prompt — a Jinja template string following the system/user/assistant message structure for chat API

LLM Node — Key Optional Parameters

The max_tokens default differs by API type: inf for chat, 16 for completion — a common exam trap.

Prompt Node — Jinja Templating

Variables in Jinja templates use {{ variable_name }} syntax. Conditional logic uses {% if %} / {% else %} / {% endif %} blocks. Loops use {% for item in list %} blocks:

{% for item in chat_history %}
user: {{ item.inputs.question }}
assistant: {{ item.outputs.answer }}
{% endfor %}
user: {{ question }}

The Prompt node's only output is a single rendered string. Wire it into an LLM node's prompt input.

Python Node — Four Requirements

1. All imports must be at the top of the snippet
2. Exactly one function decorated with the @tool decoratorThe required entry-point marker for Python node functions in Prompt Flow. Imported from promptflow and applied to exactly one function per snippet. Without this decorator, the function is not recognized as the node entry point. — the entry point
3. The @tool function must have a return statement — the return value is the node's output
4. Function parameters are automatically exposed as node Inputs (type annotations enable proper type selectors)

from promptflow import tool

@tool
def process_response(llm_output: str, max_length: int) -> str:
    return llm_output.strip()[:max_length]

For authenticated external services, declare a CustomConnection parameter:

from promptflow import tool
from promptflow.connections import CustomConnection

@tool
def call_external_api(query: str, conn: CustomConnection) -> str:
    api_key = conn.api_key
    # ... call external service ...
    return result

Wiring Nodes Together

Link nodes by referencing one node's output in another node's input:

• ${<node_name>.output} — entire output
• ${<node_name>.output.<field_name>} — a specific field of a structured output

The graph view automatically draws directed edges based on these references. Execution order is determined by dependencies, not declaration order.