Skip to content

README.md

Latest commit

 

History

History
326 lines (255 loc) · 9.19 KB

README.md

File metadata and controls

326 lines (255 loc) · 9.19 KB

LightRace

lightrace-python

PyPI version GitHub stars License

Lightweight LLM tracing SDK for Python with remote tool invocation.

Install

pip install lightrace

Quick Start

from lightrace import Lightrace, trace

lt = Lightrace(
    public_key="pk-lt-demo",
    secret_key="sk-lt-demo",
    host="http://localhost:3000",
)

# Root trace
@trace()
def run_agent(query: str):
    return search(query)

# Span
@trace(type="span")
def search(query: str) -> list:
    return ["result1", "result2"]

# Generation (LLM call)
@trace(type="generation", model="gpt-4o")
def generate(prompt: str) -> str:
    return "LLM response"

# Tool — remotely invocable from the Lightrace UI
@trace(type="tool")
def weather_lookup(city: str) -> dict:
    return {"temp": 72, "unit": "F"}

# Tool — traced but NOT remotely invocable
@trace(type="tool", invoke=False)
def read_file(path: str) -> str:
    return open(path).read()

run_agent("hello")
lt.flush()
lt.shutdown()

@trace API

@trace()                                    # Root trace
@trace(type="span")                         # Span observation
@trace(type="generation", model="gpt-4o")   # LLM generation
@trace(type="tool")                         # Tool (remotely invocable)
@trace(type="tool", invoke=False)           # Tool (trace only)

Parameters

Parameter Type Default Description
type str None "span", "generation", "tool", "chain", "event"
name str None Override name (defaults to function name)
invoke bool True For type="tool": register for remote invocation
model str None For type="generation": LLM model name
metadata dict None Static metadata attached to every call

Lightrace() Constructor

Parameter Type Default Description
tools list None LangChain tools or callables to register for dashboard re-invocation
context dict[str, (getter, setter)] None Context variables for automatic capture/restore during fork 
from lightrace import Lightrace

lt = Lightrace(
    public_key="pk-lt-demo",
    secret_key="sk-lt-demo",
    host="http://localhost:3000",
    tools=[get_weather, calculate],           # register tools in one step
    context={                                  # register context vars in one step
        "user_id": (get_user_id, set_user_id),
        "session_id": (get_session, set_session),
    },
)

Integrations

OpenAI

import openai
from lightrace import Lightrace, trace
from lightrace.integrations.openai import LightraceOpenAIInstrumentor

lt = Lightrace(
    public_key="pk-lt-demo",
    secret_key="sk-lt-demo",
    host="http://localhost:3000",
)

client = openai.OpenAI()
instrumentor = LightraceOpenAIInstrumentor(client=lt)
instrumentor.instrument(client)

@trace()
def ask_gpt():
    response = client.chat.completions.create(
        model="gpt-4o-mini",
        max_tokens=256,
        messages=[{"role": "user", "content": "What is the speed of light?"}],
    )
    return response.choices[0].message.content

ask_gpt()
lt.flush()
lt.shutdown()

Anthropic

import anthropic
from lightrace import Lightrace, trace
from lightrace.integrations.anthropic import LightraceAnthropicInstrumentor

lt = Lightrace(
    public_key="pk-lt-demo",
    secret_key="sk-lt-demo",
    host="http://localhost:3000",
)

client = anthropic.Anthropic()
instrumentor = LightraceAnthropicInstrumentor(client=lt)
instrumentor.instrument(client)

@trace()
def ask_claude():
    response = client.messages.create(
        model="claude-sonnet-4-20250514",
        max_tokens=256,
        messages=[{"role": "user", "content": "What is the capital of Mongolia?"}],
    )
    return response.content[0].text

ask_claude()
lt.flush()
lt.shutdown()

LangChain

from langchain_core.messages import HumanMessage
from langchain_openai import ChatOpenAI
from lightrace import Lightrace
from lightrace.integrations.langchain import LightraceCallbackHandler

lt = Lightrace(
    public_key="pk-lt-demo",
    secret_key="sk-lt-demo",
    host="http://localhost:3000",
)

handler = LightraceCallbackHandler(client=lt)
model = ChatOpenAI(model="gpt-4o-mini", max_tokens=256)

response = model.invoke(
    [HumanMessage(content="What is the speed of light?")],
    config={"callbacks": [handler]},
)

lt.flush()
lt.shutdown()

LangGraph Fork / Replay

Fork lets you answer "what if this tool returned something different?" by forking a LangGraph execution from any tool checkpoint and continuing with modified output.

import asyncio
from langchain_anthropic import ChatAnthropic
from langchain_core.tools import tool
from langgraph.checkpoint.memory import MemorySaver
from langgraph.prebuilt import create_react_agent
from lightrace import Lightrace
from lightrace.integrations.langchain import LightraceCallbackHandler

@tool
def get_weather(city: str) -> str:
    """Get the current weather for a city."""
    return "72F, sunny"

# Pass tools= to register them for dashboard re-invocation
lt = Lightrace(
    public_key="pk-lt-demo",
    secret_key="sk-lt-demo",
    host="http://localhost:3000",
    tools=[get_weather],
)

# Checkpointer is required for fork
agent = create_react_agent(
    ChatAnthropic(model="claude-sonnet-4-20250514"),
    [get_weather],
    checkpointer=MemorySaver(),
)

async def main():
    thread_id = "demo-thread"
    handler = LightraceCallbackHandler(
        client=lt,
        session_id=thread_id,
        trace_name="weather-agent",
        configurable={"thread_id": thread_id},
    )

    await agent.ainvoke(
        {"messages": [("user", "What's the weather in Tokyo?")]},
        config={"configurable": {"thread_id": thread_id}, "callbacks": [handler]},
    )

    # Register the graph for fork/replay from the dashboard
    lt.register_graph(agent, event_loop=asyncio.get_running_loop())
    lt.flush()

asyncio.run(main())

Requirements for fork:

  • Graph must have a checkpointer (MemorySaver, AsyncPostgresSaver, etc.)
  • Call lt.register_graph(agent) to enable fork from the dashboard
  • Pass tools=[...] to the constructor (or call lt.register_tools(...)) so tools can be re-invoked
  • Pass session_id=thread_id and configurable={"thread_id": ...} to the callback handler

Claude Agent SDK

import anyio
from claude_agent_sdk import AssistantMessage, ClaudeAgentOptions, ResultMessage, TextBlock
from lightrace import Lightrace
from lightrace.integrations.claude_agent_sdk import traced_query

lt = Lightrace(
    public_key="pk-lt-demo",
    secret_key="sk-lt-demo",
    host="http://localhost:3000",
)

async def main():
    async for message in traced_query(
        prompt="What files are in the current directory?",
        options=ClaudeAgentOptions(max_turns=3),
        client=lt,
        trace_name="file-lister",
    ):
        if isinstance(message, AssistantMessage):
            for block in message.content:
                if isinstance(block, TextBlock):
                    print(block.text)
        elif isinstance(message, ResultMessage):
            print(f"Cost: ${message.total_cost_usd:.4f}")

    lt.flush()
    lt.shutdown()

anyio.run(main)

You can also use the handler directly for more control:

from claude_agent_sdk import query
from lightrace.integrations.claude_agent_sdk import LightraceAgentHandler

handler = LightraceAgentHandler(prompt="Hello", client=lt, trace_name="my-agent")

async for message in query(prompt="Hello"):
    handler.handle(message)

Compatibility

Lightrace server also accepts traces from Langfuse Python/JS SDKs.

Related

Development

uv sync --extra dev
uv run pre-commit install
uv run pytest -s -v tests/
uv run ruff check .
uv run mypy src/lightrace

License

MIT