Skip to main content

Log custom LLM traces

note

Nothing will break if you don't log LLM traces in the correct format - data will still be logged. However, the data will not be processed or rendered in a way that is specific to LLMs.

LangSmith provides special rendering and processing for LLM traces, including token counting (assuming token counts are not available from the model provider) and token-based cost calculation. In order to make the most of this feature, you must log your LLM traces in a specific format.

note

The examples below uses the traceable decorator/wrapper to log the model run (which is the recommended approach for Python and JS/TS). However, the same idea applies if you are using the RunTree or API directly.

Chat-style models​

Using LangChain OSS or LangSmith wrappers​

If you are using OpenAI or Anthropic models, we suggest using the wrap_openai or wrap_anthropic which will automatically log traces in the format LangSmith expects.

You can also use any of the LangChain OSS chat models as a drop in for directly accessing the API of any of the providers, and this will also log LLM traces in the format LangSmith expects.

Implementing your own custom chat-model​

You may also trace your model calls using traceable or other broader tracing techniques. To properly mark the traced function as an LLM run in this case, at bare minimum you must pass run_type as "llm" like this:

from langsmith import traceable

@traceable(run_type="llm")
def chat_model(messages: list):
  ...

To make your custom LLM traces appear well-formatted in the LangSmith UI, your trace inputs and outputs must conform to a format LangSmith recognizes:

  • A list of messages in OpenAI or Anthropic format, represented as Python dictionaries or TypeScript objects.
    • Each message must contain the key role and content.
    • Messages with the "assistant" role may optionally contain tool_calls. These tool_calls may be in OpenAI format or LangChain's format.
  • An dict/object containing "messages" key with a list of messages in the above format.
    • LangSmith may use additional parameters in this input dict that match OpenAI's chat completion endpoint for rendering in the trace view, such as a list of available tools for the model to call.

Here are some examples:

# Format 1: List of messages
inputs = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "What's the weather like?"},
    {
        "role": "assistant", 
        "content": "I need to check the weather for you.",
        "tool_calls": [
            {
                "id": "call_123",
                "type": "function",
                "function": {
                    "name": "get_weather",
                    "arguments": '{"location": "current"}'
                }
            }
        ]
    }
]

@traceable(run_type="llm")
def chat_model(messages: list):
    ...

chat_model(inputs)

The output is accepted in any of the following formats:

  • A dictionary/object that contains the key choices with a value that is a list of dictionaries/objects. Each dictionary/object must contain the key message, which maps to a message object with the keys role and content.
  • A dictionary/object that contains the key message with a value that is a message object with the keys role and content.
  • A tuple/array of two elements, where the first element is the role and the second element is the content.
  • A dictionary/object that contains the key role and content.

Here are some examples:

from langsmith import traceable

@traceable(run_type="llm")
def chat_model_choices(messages):
    # Your model logic here
    return {
        "choices": [
            {
                "message": {
                    "role": "assistant",
                    "content": "Sure, what time would you like to book the table for?"
                }
            }
        ]
    }

# Usage
inputs = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "I'd like to book a table for two."}
]
chat_model_choices(inputs)

You can also provide the following metadata fields to help LangSmith identify the model - which if recognized, LangSmith will use to automatically calculate costs. To learn more about how to use the metadata fields, see this guide.

  • ls_provider: The provider of the model, eg "openai", "anthropic", etc.
  • ls_model_name: The name of the model, eg "gpt-4o-mini", "claude-3-opus-20240307", etc.
from langsmith import traceable

inputs = [
  {"role": "system", "content": "You are a helpful assistant."},
  {"role": "user", "content": "I'd like to book a table for two."},
]

output = {
  "choices": [
      {
          "message": {
              "role": "assistant",
              "content": "Sure, what time would you like to book the table for?"
          }
      }
  ]
}

# Can also use one of:
# output = {
#     "message": {
#         "role": "assistant",
#         "content": "Sure, what time would you like to book the table for?"
#     }
# }
#
# output = {
#     "role": "assistant",
#     "content": "Sure, what time would you like to book the table for?"
# }
#
# output = ["assistant", "Sure, what time would you like to book the table for?"]

@traceable(
  run_type="llm",
  metadata={"ls_provider": "my_provider", "ls_model_name": "my_model"}
)
def chat_model(messages: list):
  return output

chat_model(inputs)

The above code will log the following trace:

If you implement a custom streaming chat_model, you can "reduce" the outputs into the same format as the non-streaming version. This is currently only supported in Python.

def _reduce_chunks(chunks: list):
    all_text = "".join([chunk["choices"][0]["message"]["content"] for chunk in chunks])
    return {"choices": [{"message": {"content": all_text, "role": "assistant"}}]}

@traceable(
    run_type="llm",
    reduce_fn=_reduce_chunks,
    metadata={"ls_provider": "my_provider", "ls_model_name": "my_model"}
)
def my_streaming_chat_model(messages: list):
    for chunk in ["Hello, " + messages[1]["content"]]:
        yield {
            "choices": [
                {
                    "message": {
                        "content": chunk,
                        "role": "assistant",
                    }
                }
            ]
        }

list(
    my_streaming_chat_model(
        [
            {"role": "system", "content": "You are a helpful assistant. Please greet the user."},
            {"role": "user", "content": "polly the parrot"},
        ],
    )
)
tip

If ls_model_name is not present in extra.metadata, other fields might be used from the extra.metadata for estimating token counts. The following fields are used in the order of precedence:

  1. metadata.ls_model_name
  2. inputs.model
  3. inputs.model_name

Provide token and cost information​

By default, LangSmith uses tiktoken to count tokens, utilizing a best guess at the model's tokenizer based on the ls_model_name provided. It also calculates costs automatically by using the model pricing table. To learn how LangSmith calculates token-based costs, see this guide.

However, many models already include exact token counts as part of the response. If you have this information, you can override the default token calculation in LangSmith in one of two ways:

  1. Extract usage within your traced function and set a usage_metadata field on the run's metadata.
  2. Return a usage_metadata field in your traced function outputs.

In both cases, the usage metadata you send should contain a subset of the following LangSmith-recognized fields:

Setting usage metadata

You cannot set any fields other than the ones listed below. You do not need to include all fields.

class UsageMetadata(TypedDict, total=False):
    input_tokens: int
    """The number of tokens used for the prompt."""
    output_tokens: int
    """The number of tokens generated as output."""
    total_tokens: int
    """The total number of tokens used."""
    input_token_details: dict[str, float]
    """The details of the input tokens."""
    output_token_details: dict[str, float]
    """The details of the output tokens."""
    input_cost: float
    """The cost of the input tokens."""
    output_cost: float
    """The cost of the output tokens."""
    total_cost: float
    """The total cost of the tokens."""
    input_cost_details: dict[str, float]
    """The cost details of the input tokens."""
    output_cost_details: dict[str, float]
    """The cost details of the output tokens."""

Note that the usage data can also include cost information, in case you do not want to rely on LangSmith's token-based cost formula. This is useful for models with pricing that is not linear by token type.

Setting run metadata​

You can modify the current run's metadata with usage information within your traced function. The advantage of this approach is that you do not need to change your traced function's runtime outputs. Here's an example:

Dependencies

Requires langsmith>=0.3.43 (Python) and langsmith>=0.3.30 (JS/TS).

from langsmith import traceable, get_current_run_tree

inputs = [
  {"role": "system", "content": "You are a helpful assistant."},
  {"role": "user", "content": "I'd like to book a table for two."},
]

@traceable(
  run_type="llm",
  metadata={"ls_provider": "my_provider", "ls_model_name": "my_model"}
)
def chat_model(messages: list):
  llm_output = {
      "choices": [
          {
              "message": {
                  "role": "assistant",
                  "content": "Sure, what time would you like to book the table for?"
              }
          }
      ],
      "usage_metadata": {
          "input_tokens": 27,
          "output_tokens": 13,
          "total_tokens": 40,
          "input_token_details": {"cache_read": 10},
          # If you wanted to specify costs:
          # "input_cost": 1.1e-6,
          # "input_cost_details": {"cache_read": 2.3e-7},
          # "output_cost": 5.0e-6,
      },
  }
  run = get_current_run_tree()
  run.set(usage_metadata=llm_output["usage_metadata"])
  return llm_output["choices"][0]["message"]

chat_model(inputs)

Setting run outputs​

You can add a usage_metadata key to the function's response to set manual token counts and costs.

from langsmith import traceable

inputs = [
  {"role": "system", "content": "You are a helpful assistant."},
  {"role": "user", "content": "I'd like to book a table for two."},
]

output = {
  "choices": [
      {
          "message": {
              "role": "assistant",
              "content": "Sure, what time would you like to book the table for?"
          }
      }
  ],
  "usage_metadata": {
      "input_tokens": 27,
      "output_tokens": 13,
      "total_tokens": 40,
      "input_token_details": {"cache_read": 10},
      # If you wanted to specify costs:
      # "input_cost": 1.1e-6,
      # "input_cost_details": {"cache_read": 2.3e-7},
      # "output_cost": 5.0e-6,
  },
}

@traceable(
  run_type="llm",
  metadata={"ls_provider": "my_provider", "ls_model_name": "my_model"}
)
def chat_model(messages: list):
  return output

chat_model(inputs)

Instruct-style models​

For instruct-style models (string in, string out), your inputs must contain a key prompt with a string value. Other inputs are also permitted. The output must return an object that, when serialized, contains the key choices with a list of dictionaries/objects. Each must contain the key text with a string value. The same rules for metadata and usage_metadata apply as for chat-style models.

@traceable(
  run_type="llm",
  metadata={"ls_provider": "my_provider", "ls_model_name": "my_model"}
)
def hello_llm(prompt: str):
  return {
      "choices": [
          {"text": "Hello, " + prompt}
      ],
      "usage_metadata": {
          "input_tokens": 4,
          "output_tokens": 5,
          "total_tokens": 9,
      },
  }

hello_llm("polly the parrot\n")

The above code will log the following trace:

Was this page helpful?


You can leave detailed feedback on GitHub.