Viewing trace details and latency in LangChain

Introduction

Seeing trace details and latency helps you understand how your LangChain app works step-by-step and how fast each part runs.

You want to check which part of your chain is slow.

You need to debug why your app is not giving expected results.

You want to improve performance by finding bottlenecks.

You want to see the inputs and outputs of each step clearly.

You want to log detailed info for monitoring your app.

Syntax

LangChain

from langchain.callbacks import get_openai_callback

with get_openai_callback() as cb:
    result = chain.run("Your input here")
    print(cb)

# For detailed tracing, use tracing tools or callbacks that show step info and timing.

The get_openai_callback() helps measure token usage and cost for OpenAI calls.

For full trace details, LangChain supports callback handlers that log each step's input, output, and time.

Examples

This shows token usage and cost for the OpenAI call inside the chain.

LangChain

from langchain.callbacks import get_openai_callback

with get_openai_callback() as cb:
    result = chain.run("Hello")
    print(cb)

This example uses a tracer callback to get detailed trace info of the chain execution.

LangChain

from langchain.callbacks.tracers import LangChainTracer
from langchain.chains import LLMChain

tracer = LangChainTracer()
chain = LLMChain(llm=llm, prompt=prompt, callbacks=[tracer])

result = chain.run("Hello")
print(tracer.get_trace())

Sample Program

This program runs a simple LangChain chain that says hello to a name. It uses a tracer callback to capture detailed trace info including inputs, outputs, and latency. It prints the final result and the trace details.

LangChain

from langchain.llms import OpenAI
from langchain.chains import LLMChain
from langchain.prompts import PromptTemplate
from langchain.callbacks.tracers import LangChainTracer

# Create a simple prompt
prompt = PromptTemplate(template="Say hello to {name}", input_variables=["name"])

# Initialize OpenAI LLM
llm = OpenAI(temperature=0)

# Create tracer to capture trace details
tracer = LangChainTracer()

# Create chain with tracer callback
chain = LLMChain(llm=llm, prompt=prompt, callbacks=[tracer])

# Run chain
result = chain.run({"name": "Alice"})

# Print result and trace details
print("Result:", result)
print("Trace details:", tracer.get_trace())

Important Notes

Trace details include inputs, outputs, start and end times, and latency for each step.

Latency helps find slow parts to optimize.

Use callbacks to customize what trace info you want to collect.

Summary

Viewing trace details helps you see what happens inside your LangChain app.

Latency shows how long each step takes, helping find slow parts.

Use built-in callbacks like LangChainTracer to get detailed trace info easily.

Practice

1. What is the main purpose of viewing trace details in a LangChain application?

easy

A. To change the output format of the application

B. To understand the internal steps and flow of the application

C. To increase the speed of the application automatically

D. To add new features without coding

Solution

1. Step 1: Understand what trace details represent

   Trace details show the internal steps and flow of the LangChain app during execution.

2. Step 2: Identify the purpose of viewing these details

   Viewing trace details helps developers see what happens inside, making debugging and optimization easier.

3. Final Answer:

   To understand the internal steps and flow of the application -> Option B

4. Quick Check:

   Trace details = internal flow insight [OK]

Hint: Trace details show what happens inside your app [OK]

Common Mistakes:

• Thinking trace changes output format
• Believing trace speeds up app automatically
• Confusing trace with adding features

2. Which of the following is the correct way to enable LangChainTracer to view trace details?

easy

A. from langchain.callbacks import tracer\ntracer = tracer()

B. import LangChainTracer from langchain.callbacks\ntracer = LangChainTracer()

C. from langchain import LangChainTracer\ntracer = LangChainTracer()

D. from langchain.callbacks import LangChainTracer\ntracer = LangChainTracer()

Solution

1. Step 1: Recall the correct import path for LangChainTracer

   LangChainTracer is imported from langchain.callbacks module.

2. Step 2: Check the correct instantiation syntax

   The correct way is to import LangChainTracer and then create an instance with LangChainTracer().

3. Final Answer:

   from langchain.callbacks import LangChainTracer\ntracer = LangChainTracer() -> Option D

4. Quick Check:

   Correct import and instantiation = from langchain.callbacks import LangChainTracer\ntracer = LangChainTracer() [OK]

Hint: Import from langchain.callbacks and instantiate with () [OK]

Common Mistakes:

• Wrong import path
• Using incorrect import syntax
• Calling a non-existent function

3. Given this code snippet using LangChainTracer, what will be the output regarding latency?

from langchain.callbacks import LangChainTracer
tracer = LangChainTracer()

with tracer:
    result = chain.run("Hello")

print(tracer.get_trace())

medium

A. A detailed trace including each step's latency in milliseconds

B. Only the final output without any timing information

C. An error because get_trace() is not a valid method

D. An empty trace because tracer was not started

Solution

1. Step 1: Understand what LangChainTracer does inside a with block

   LangChainTracer collects detailed trace info including latency for each step during the with block execution.

2. Step 2: Check the output of get_trace()

   get_trace() returns the collected trace data, which includes latency details for each step.

3. Final Answer:

   A detailed trace including each step's latency in milliseconds -> Option A

4. Quick Check:

   Tracer with block + get_trace() = detailed latency trace [OK]

Hint: Tracer inside with block captures latency info [OK]

Common Mistakes:

• Assuming get_trace() does not exist
• Expecting output without timing info
• Forgetting to use with block for tracing

4. You added LangChainTracer but see no trace output after running your chain. What is the most likely cause?

medium

A. The chain.run() method does not support tracing

B. You did not import LangChainTracer correctly

C. You forgot to wrap the chain execution inside the tracer's with block

D. You need to call tracer.start() before running the chain

Solution

1. Step 1: Check how LangChainTracer collects trace data

   LangChainTracer collects trace only when chain execution is inside its with block context.

2. Step 2: Identify the missing step causing no trace output

   If chain.run() is outside the with block, no trace is recorded, so output is empty.

3. Final Answer:

   You forgot to wrap the chain execution inside the tracer's with block -> Option C

4. Quick Check:

   Tracing requires with block wrapping chain run [OK]

Hint: Always run chain inside tracer's with block [OK]

Common Mistakes:

• Assuming import error causes no trace
• Thinking chain.run() disables tracing
• Trying to call non-existent start() method

5. You want to find the slowest step in your LangChain app using LangChainTracer. Which approach correctly identifies it?

hard

A. Extract latency from each trace step and find the maximum value

B. Check the total runtime of the app without step details

C. Use print statements inside the chain to guess slow parts

D. Restart the app multiple times to see when it feels slow

Solution

1. Step 1: Understand what latency data LangChainTracer provides

   LangChainTracer records latency for each individual step in the trace details.

2. Step 2: Identify how to find the slowest step

   By extracting latency values from each step and comparing them, you can find the maximum latency which indicates the slowest step.

3. Final Answer:

   Extract latency from each trace step and find the maximum value -> Option A

4. Quick Check:

   Find max latency in trace steps = slowest step [OK]

Hint: Find max latency value in trace steps to spot slowest [OK]

Common Mistakes:

• Ignoring step-level latency and checking only total time
• Using print statements instead of trace data
• Guessing speed without data