LangChain

The Maritaca API is compatible with the OpenAI Chat Completions API. In LangChain, use the ChatOpenAI class from the langchain-openai package and set Maritaca's endpoint as the base_url.

Migrating from the old integration

The ChatMaritalk class from langchain-community is a legacy integration. Examples that use sabia-2-small or sabia-2-medium are also outdated because those models have been discontinued. Use ChatOpenAI with a current model as shown below.

Installation

Install LangChain's OpenAI integration:

pip install -U langchain-openai

Set your API key in an environment variable. You can create a key in the Maritaca platform.

export MARITACA_API_KEY="your-key-here"

Regular call

import os

from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    model="sabia-4",
    api_key=os.environ["MARITACA_API_KEY"],
    base_url="https://chat.maritaca.ai/api",
)

response = llm.invoke(
    [
        ("system", "Answer directly and briefly."),
        ("human", "What is 25 + 27?"),
    ]
)

print(response.content)

Both models use the same configuration:

Model	When to use it
sabia-4	General use with a good balance of quality, speed, and cost.
sabia-4-thinking	Complex tasks that benefit from reasoning before the answer.

To use the reasoning model, change only the model name:

llm = ChatOpenAI(
    model="sabia-4-thinking",
    api_key=os.environ["MARITACA_API_KEY"],
    base_url="https://chat.maritaca.ai/api",
)

Tools (function calling)

Use bind_tools to send tool definitions to the model. The model requests a tool call, but your application is responsible for executing it and returning the result in a ToolMessage.

The example below forces the first tool call only to make the demonstration reproducible. In a regular application, omit tool_choice and let the model decide when to use the tool.

import os

from langchain_core.messages import HumanMessage, ToolMessage
from langchain_core.tools import tool
from langchain_openai import ChatOpenAI


@tool
def multiply(a: int, b: int) -> int:
    """Multiply two integers."""
    return a * b


llm = ChatOpenAI(
    model="sabia-4-thinking",
    api_key=os.environ["MARITACA_API_KEY"],
    base_url="https://chat.maritaca.ai/api",
)

question = "Use the multiply tool to calculate 6 times 7."
messages = [HumanMessage(content=question)]

llm_with_forced_tool = llm.bind_tools([multiply], tool_choice="multiply")
tool_request = llm_with_forced_tool.invoke(messages)
messages.append(tool_request)

if not tool_request.tool_calls:
    raise RuntimeError("The model did not request a tool")

tool_call = tool_request.tool_calls[0]
tool_result = multiply.invoke(tool_call["args"])
messages.append(
    ToolMessage(content=str(tool_result), tool_call_id=tool_call["id"])
)

final_response = llm.bind_tools([multiply]).invoke(messages)
print(final_response.content)  # 42

This flow works with both sabia-4 and sabia-4-thinking. For more information about messages, chains, and agents, see the ChatOpenAI documentation.