Copy page

Build with Hyphen

LangChain & LangGraph

LangChain is the largest LLM framework ecosystem in Python. LangGraph is its graph-based agent runtime. Both reach Hyphen through ChatOpenAI with two changed kwargs.

Code
pip install langchain langchain-openai langgraph
export HYPHEN_API_KEY="sk-..."

Setup guide: this config comes from LangChain's official documentation and has not been run end to end against the gateway. Corrections to support@hyphen-solution.com.

The model

Code
import os
from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    model="minimax-m3",
    base_url="https://api.hyphen-solution.com/v1",
    api_key=os.environ["HYPHEN_API_KEY"],
    max_tokens=2000,
    temperature=0,
)

print(llm.invoke("Explain a Bloom filter in three sentences.").content)

base_url, api_key and model are the canonical kwarg names. openai_api_base, openai_api_key and model_name still work as aliases from older tutorials, but use the short ones.

Streaming

Code
for chunk in llm.stream("Count to ten slowly."):
    print(chunk.content, end="", flush=True)

Tools

Code
from langchain_core.tools import tool


@tool
def get_weather(city: str) -> str:
    """Get the current weather for a city."""
    return f"It is 31C and storming in {city}."


bound = llm.bind_tools([get_weather])
resp = bound.invoke("What is the weather in Kuala Lumpur?")
print(resp.tool_calls)

A LangGraph agent

Code
import os
from langchain_openai import ChatOpenAI
from langchain_core.tools import tool
from langchain.agents import create_agent

llm = ChatOpenAI(
    model="minimax-m3",
    base_url="https://api.hyphen-solution.com/v1",
    api_key=os.environ["HYPHEN_API_KEY"],
    max_tokens=4000,
    temperature=0,
)


@tool
def get_weather(city: str) -> str:
    """Get the current weather for a city."""
    return f"It is 31C and storming in {city}."


agent = create_agent(
    model=llm,
    tools=[get_weather],
    system_prompt="You are a helpful assistant. Use the tools you have.",
)

result = agent.invoke(
    {"messages": [{"role": "user", "content": "What is the weather in Kuala Lumpur?"}]}
)
print(result["messages"][-1].content)

Structured output

with_structured_output defaults to a function-calling strategy, which is what you want here. MiniMax supports tool calling but not response_format, so do not pass method="json_mode" or method="json_schema".

Code
from pydantic import BaseModel, Field


class Issue(BaseModel):
    """A structured bug report."""

    title: str = Field(description="One-line summary")
    severity: str = Field(description="low, medium, high or critical")
    component: str


structured = llm.with_structured_output(Issue)  # method="function_calling" is the default
print(structured.invoke("The CSV exporter drops the last row on 2.4.1."))

Gotchas

• max_tokens of 2000 or more. 4000 for agents. The M-series reason before answering and a tight cap returns an empty AIMessage. See Choosing a model.
• No OpenAIEmbeddings against Hyphen. The catalog has no embedding models. Point your vector store somewhere else.
• Doc links moved. The API reference now lives at reference.langchain.com.
• 429 means your monthly budget is spent. See Handling the 429 cap.

Related

• LangChain agents docs
• ChatOpenAI reference
• Structured JSON output: why tool calling and not JSON mode.
• Choosing a model: which model for which job.