Anthropic with LangChain (System Message)

[oliverbj]

Hi.

In the LangChain docs, it says this about Anthropic's system message:

No System Messages

Anthropic models are not trained on the concept of a "system message". We have worked with the Anthropic team to handle them somewhat appropriately (a Human message with an admin tag) but this is largely a hack and it is recommended that you do not use system messages.

However, since version 2.1, Claude now support system prompts. Can this be updated in LangChain as well?

[openerror]

You can specify system messages via model_kwargs when creating the LLM or Chatbot object. Like this: ChatAnthropic(..., model_kwargs=dict(system='Your system prompt')) here.

[superchordate]

It seems the correct method is to include a system message in the initial message invokation:

const agent = createReactAgent({
    llm: new ChatAnthropic({
      modelName: model,
      maxTokens: 1024,
      temperature: 0.3,
    }),
    tools: [],
    checkpointSaver: new MemorySaver(),
  });
  const agentState = await agent.invoke(
    {
      messages: [
        new SystemMessage(
          "You are a testing agent that always response with 'test' regardless of what the user asks."
        ),
        new HumanMessage("What is your role?"),
      ],
    },
    { configurable: { thread_id: "42" } }
  );

This will respond with "test".