LangChain LangGraph
Module 1 Module 2 Module 3
Module 3 · Middleware and HITL
1 3.2 Managing Msgs
2 3.3 HITL
3 3.4 Dyn Models
4 3.4 Dyn Prompts
5 3.4 Dyn Tools
6 3.5 Email Agent
SUM Module Summary
Home›LangChain›Module 3 · Middleware and HITL›3.4 Dyn Prompts
📓 Notebook 📖 Explained
LANGCHAIN MODULE 3 · LESSON 3.4b

Dynamic Prompts
dynamic_prompt · Runtime Context

This section dynamically generates system prompts based on the context at run time, allowing the same Agent to respond in different languages ​​to different User users.

1 Uses for dynamic prompts

System prompt is not necessarily a fixed string. It can be dynamically generated based on user language, role, permissions, region or business context. In this lesson, UseUse user language will be used as a minimal example.

input contextGenerated system prompt
EnglishYou are a helpful assistant.
IrishYou are a helpful assistant. only respond in Irish.
SpanishYou are a helpful assistant. only respond in Spanish.

2 Define LanguageContext

Run-time context Use dataclass expression. When calling Use Agent, pass different user_language, and the middleware can generate different prompts.

Python
from dataclasses import dataclass
from langchain.agents.middleware import dynamic_prompt, ModelRequest

@dataclass
class LanguageContext:
    user_language: str = "English"
Connection to Module 2

Module 2 has learned about runtime context; in this section, Put context Use is generated at the prompt instead of being read by the tool.

3 Use @dynamic_prompt to generate the system prompt

The @dynamic_prompt function receives ModelRequest and can access request.runtime.context. The return value will become the system prompt for this model adjustment.

Python
@dynamic_prompt
def user_language_prompt(request: ModelRequest) -> str:
    """Generate system prompt based on user role."""
    user_language = request.runtime.context.user_language
    base_prompt = "You are a helpful assistant."

    if user_language != "English":
        return f"{base_prompt} only respond in {user_language}."
    elif user_language == "English":
        return base_prompt
The return value is a prompt

This middleware does not directly answer the Use user, but generates system messages for the model to call Use. The real answer still lies with the model.

4 Create Agent

Declare context_schema and middleware when creating Agent. Each subsequent invoke can pass a different context.

Python
agent = create_agent(
    model=deepseek_model(),
    context_schema=LanguageContext,
    middleware=[user_language_prompt],
)
No need for multiple Agents

Don't Use to copy one Agent for each language. When an Agent adds a dynamic prompt, it can switch behaviors according to context.

5 Call the same Agent in different languages

Notebook inputs Irish, Spanish, and French respectively. The questions are the same, but the system prompts are different, so the answer language will also change.

Python
response = agent.invoke(
    {"messages": [HumanMessage(content="Hello, how are you?")]},
    context=LanguageContext(user_language="Irish"),
)
print(response["messages"][-1].content)

response = agent.invoke(
    {"messages": [HumanMessage(content="Hello, how are you?")]},
    context=LanguageContext(user_language="Spanish"),
)
print(response["messages"][-1].content)

response = agent.invoke(
    {"messages": [HumanMessage(content="Hello, how are you?")]},
    context=LanguageContext(user_language="French"),
)
print(response["messages"][-1].content)
Extensible pattern

Language is just an example. The same model can generate different rules according to user roles, different terms according to product lines, and different compliance prompts according to regions.

6 Lesson takeaway