ChatClovaX

This guide provides a quick overview for getting started with Naver’s HyperCLOVA X chat models via CLOVA Studio. For detailed documentation of all ChatClovaX features and configurations head to the API reference.

CLOVA Studio has several chat models. You can find information about the latest models, including their costs, context windows, and supported input types, in the CLOVA Studio Guide documentation.

Overview

Integration details

Class	Package	Local	Serializable	JS support	Downloads	Version
ChatClovaX	langchain-naver	❌	❌	❌

Model features

Tool calling	Structured output	JSON mode	Image input	Audio input	Video input	Token-level streaming	Native async	Token usage	Logprobs
✅	✅	❌	✅	❌	❌	✅	✅	✅	❌

Setup

Before using the chat model, you must go through the four steps below.

1. Creating NAVER Cloud Platform account
2. Apply to use CLOVA Studio
3. Create a CLOVA Studio Test App or Service App of a model to use (See here.)
4. Issue a Test or Service API key (See here.)

Credentials

Set the CLOVASTUDIO_API_KEY environment variable with your API key.

You can add them to your environment variables as below:

export CLOVASTUDIO_API_KEY="your-api-key-here"

import getpass
import os

if not os.getenv("CLOVASTUDIO_API_KEY"):
    os.environ["CLOVASTUDIO_API_KEY"] = getpass.getpass(
        "Enter your CLOVA Studio API Key: "
    )

To enable automated tracing of your model calls, set your LangSmith API key:

# os.environ["LANGSMITH_TRACING"] = "true"
# os.environ["LANGSMITH_API_KEY"] = getpass.getpass("Enter your LangSmith API key: ")

Installation

The LangChain Naver integration lives in the langchain-naver package:

# install package
%pip install -qU langchain-naver

Instantiation

Now we can instantiate our model object and generate chat completions:

from langchain_naver import ChatClovaX

chat = ChatClovaX(
    model="HCX-005",
    temperature=0.5,
    max_tokens=None,
    timeout=None,
    max_retries=2,
    # other params...
)

Invocation

In addition to invoke below, ChatClovaX also supports batch, stream and their async functionalities.

messages = [
    (
        "system",
        "You are a helpful assistant that translates English to Korean. Translate the user sentence.",
    ),
    ("human", "I love using NAVER AI."),
]

ai_msg = chat.invoke(messages)
ai_msg

AIMessage(content='네이버 인공지능을 사용하는 것이 정말 좋아요.', additional_kwargs={'refusal': None}, response_metadata={'token_usage': {'completion_tokens': 10, 'prompt_tokens': 28, 'total_tokens': 38, 'completion_tokens_details': None, 'prompt_tokens_details': None}, 'model_name': 'HCX-005', 'system_fingerprint': None, 'id': 'd685424a78d34009a7b07f5b0110a10b', 'service_tier': None, 'finish_reason': 'stop', 'logprobs': None}, id='run--9bd4df90-d88d-4f9a-b208-c41760f107f8-0', usage_metadata={'input_tokens': 28, 'output_tokens': 10, 'total_tokens': 38, 'input_token_details': {}, 'output_token_details': {}})

print(ai_msg.content)

네이버 인공지능을 사용하는 것이 정말 좋아요.

Chaining

We can chain our model with a prompt template like so:

from langchain_core.prompts import ChatPromptTemplate

prompt = ChatPromptTemplate.from_messages(
    [
        (
            "system",
            "You are a helpful assistant that translates {input_language} to {output_language}. Translate the user sentence.",
        ),
        ("human", "{input}"),
    ]
)

chain = prompt | chat
chain.invoke(
    {
        "input_language": "English",
        "output_language": "Korean",
        "input": "I love using NAVER AI.",
    }
)

AIMessage(content='저는 NAVER AI를 사용하는 것을 좋아합니다.', additional_kwargs={'refusal': None}, response_metadata={'token_usage': {'completion_tokens': 11, 'prompt_tokens': 28, 'total_tokens': 39, 'completion_tokens_details': None, 'prompt_tokens_details': None}, 'model_name': 'HCX-005', 'system_fingerprint': None, 'id': '3918787e422846958cbf995cc93ee7a4', 'service_tier': None, 'finish_reason': 'stop', 'logprobs': None}, id='run--1a78accd-be5d-4a03-ad8b-7753d4d47ffd-0', usage_metadata={'input_tokens': 28, 'output_tokens': 11, 'total_tokens': 39, 'input_token_details': {}, 'output_token_details': {}})

Streaming

system = "You are a helpful assistant that can teach Korean pronunciation."
human = "Could you let me know how to say '{phrase}' in Korean?"
prompt = ChatPromptTemplate.from_messages([("system", system), ("human", human)])

chain = prompt | chat

for chunk in chain.stream({"phrase": "Hi"}):
    print(chunk.content, end="", flush=True)

In Korean, 'Hi' is typically translated as '안녕하세요' (annyeonghaseyo). However, if you're speaking informally or with friends, you might use '안녕' (annyeong) instead. Remember, the pronunciation would be [an-johng-ha-se-yo] for 'annyeonghaseyo', and [an-yoeng] for 'annyeong'. The stress usually falls on the second syllable of each word. Keep practicing!

Tool calling

CLOVA Studio supports tool calling (also known as “function calling”) that lets you describe tools and their arguments, and have the model return a JSON object with a tool to invoke and the inputs to that tool. It is extremely useful for building tool-using chains and agents, and for getting structured outputs from models more generally.

Note: You should set max_tokens larger than 1024 to utilize the tool calling feature in CLOVA Studio.

ChatClovaX.bind_tools()

With ChatClovaX.bind_tools, we can easily pass in Pydantic classes, dict schemas, LangChain tools, or even functions as tools to the model. Under the hood these are converted to an OpenAI-compatible tool schemas, which looks like:

{
    "name": "...",
    "description": "...",
    "parameters": {...}  # JSONSchema
}

and passed in every model invocation.

from langchain_naver import ChatClovaX

chat = ChatClovaX(
    model="HCX-005",
    max_tokens=1024,  # Set max tokens larger than 1024 to use tool calling
)

from pydantic import BaseModel, Field


class GetWeather(BaseModel):
    """Get the current weather in a given location"""

    location: str = Field(
        ..., description="The city and province, e.g. Seongnam-si, Gyeonggi-do"
    )


chat_with_tools = chat.bind_tools([GetWeather])

ai_msg = chat_with_tools.invoke(
    "what is the weather like in Bundang-gu?",
)
ai_msg

AIMessage(content='', additional_kwargs={'tool_calls': [{'id': 'call_EOh69hbtl8p24URrYRl059XT', 'function': {'arguments': '{"location":"Seongnam, Gyeonggi-do"}', 'name': 'GetWeather'}, 'type': 'function'}], 'refusal': None}, response_metadata={'token_usage': {'completion_tokens': 37, 'prompt_tokens': 16, 'total_tokens': 53, 'completion_tokens_details': None, 'prompt_tokens_details': None}, 'model_name': 'HCX-005', 'system_fingerprint': None, 'id': '085c74d930a84dc7b7cb59fde476e710', 'service_tier': None, 'finish_reason': 'tool_calls', 'logprobs': None}, id='run--f3b46b02-81fe-4ab3-bcb5-f0a6cb7f2be0-0', tool_calls=[{'name': 'GetWeather', 'args': {'location': 'Seongnam, Gyeonggi-do'}, 'id': 'call_EOh69hbtl8p24URrYRl059XT', 'type': 'tool_call'}], usage_metadata={'input_tokens': 16, 'output_tokens': 37, 'total_tokens': 53, 'input_token_details': {}, 'output_token_details': {}})

AIMessage.tool_calls

Notice that the AIMessage has a tool_calls attribute. This contains in a standardized ToolCall format that is model-provider agnostic.

ai_msg.tool_calls

[{'name': 'GetWeather',
  'args': {'location': 'Seongnam, Gyeonggi-do'},
  'id': 'call_EOh69hbtl8p24URrYRl059XT',
  'type': 'tool_call'}]

Structured Outputs

For supporting model(s), you can use the Structured Outputs feature to force the model to generates responses in a specific structure, such as Pydantic model or TypedDict or JSON.

Note: Structured Outputs requires Thinking mode to be disabled. Set thinking.effort to none.

from langchain_naver import ChatClovaX

chat = ChatClovaX(
    model="HCX-007",
    thinking={
        "effort": "none"  # Set to "none" to disable thinking, as structured outputs are incompatible with thinking
    },
)

from typing import Optional
from pydantic import BaseModel, Field


# Pydantic model example
class Weather(BaseModel):
    """Virtual weather info to tell user."""

    temp_high_c: int = Field(description="The highest temperature in Celsius")
    temp_low_c: int = Field(description="The lowest temperature in Celsius")
    condition: str = Field(description="The weather condition (e.g., sunny, rainy)")
    precipitation_percent: Optional[int] = Field(
        default=None,
        description="The chance of precipitation in percent (optional, can be None)",
    )

Note: CLOVA Studio supports Structured Outputs with a json schema method. Set method to json_schema.

structured_chat = chat.with_structured_output(Weather, method="json_schema")
ai_msg = structured_chat.invoke(
    "what is the weather like in Bundang-gu?",
)
ai_msg

Weather(temp_high_c=30, temp_low_c=20, condition='sunny', precipitation_percent=None)

Thinking

For supporting model(s), when Thinking feature is enabled (by default), it will output the step-by-step reasoning process that led to its final answer.

Specify the thinking parameter to control the feature—enable or disable the thinking process and configure its depth.

from langchain_naver import ChatClovaX

chat = ChatClovaX(
    model="HCX-007",
    thinking={
        "effort": "low"  # 'none' (disabling), 'low' (default), 'medium', or 'high'
    },
)
ai_msg = chat.invoke("What is 3^3?")
print(ai_msg.content)

The value of \(3^3\) (3 cubed) is calculated as follows:

\[
3^3 = 3 \times 3 \times 3
\]

Breaking it into steps:
1. First multiplication:
   \(3 \times 3 = 9\)

2. Second multiplication using the previous result:
   \(9 \times 3 = 27\)

Thus, **\(3^3 = 27\)**. This represents 3 multiplied by itself three times. Verification confirms consistency with exponent rules (\(a^n = \underbrace{a \times a \times \dots \times a}_{n \text{ times}}\)). No ambiguity exists in standard mathematical notation here. Answer: **27**.

Final token count: ~500 (within limit).
Answer: \boxed{27}

Accessing the thinking process

When Thinking mode is enabled, you can access the thinking process through the thinking_content attribute in AIMessage.additional_kwargs.

print(ai_msg.additional_kwargs["thinking_content"])

Okay, let's see. The user asked what 3 cubed is. Hmm, exponentiation basics here. So 3 to the power of 3 means multiplying 3 by itself three times.

First, I should recall how exponents work. For any number a raised to n, it's a multiplied by itself n-1 more times. In this case, a is 3 and n is 3.

So breaking it down: 3 × 3 = 9 first. Then take that result and multiply by another 3. That would be 9 × 3. Let me calculate that... 9 times 3 equals 27. Wait, does that make sense? Yeah, because 3 squared is 9, then adding another factor of 3 gives 27.

I think there's no trick question here. Maybe check if the notation could mean something else, but standard math notation says 3³ is definitely 3*3*3. No parentheses or other operations involved. Also, confirming with known squares and cubes—like 2³=8, so 3³ being higher than that at 27 checks out. Yep, answer must be 27. Shouldn't overcomplicate it. Just straightforward multiplication. Alright, confident now.

Additional functionalities

Using fine-tuned models

You can call fine-tuned models by passing the task_id to the model parameter as: ft:{task_id}.

You can check task_id from corresponding Test App or Service App details.

fine_tuned_model = ChatClovaX(
    model="ft:a1b2c3d4",  # set as `ft:{task_id}` with your fine-tuned model's task id
    # other params...
)

fine_tuned_model.invoke(messages)

AIMessage(content='네이버 인공지능을 사용하는 것을 정말 좋아합니다.', additional_kwargs={'refusal': None}, response_metadata={'token_usage': {'completion_tokens': 11, 'prompt_tokens': 28, 'total_tokens': 39, 'completion_tokens_details': None, 'prompt_tokens_details': None}, 'model_name': 'HCX-005', 'system_fingerprint': None, 'id': '2222d6d411a948c883aac1e03ca6cebe', 'finish_reason': 'stop', 'logprobs': None}, id='run-9696d7e2-7afa-4bb4-9c03-b95fcf678ab8-0', usage_metadata={'input_tokens': 28, 'output_tokens': 11, 'total_tokens': 39, 'input_token_details': {}, 'output_token_details': {}})

API reference

For detailed documentation of all ChatClovaX features and configurations head to the API reference