LangSmith evaluators in LangChain - Step-by-Step Execution

Concept Flow - LangSmith evaluators

Input: Model Output + Reference

↓

Evaluator Receives Data

↓

Run Evaluation Logic

↓

Generate Score or Feedback

↓

Return Evaluation Result

The evaluator takes the model output and reference, runs evaluation logic, and returns a score or feedback.

Execution Sample

LangChain

from langchain.evaluation import StringEvaluator

eval = StringEvaluator()
result = eval.evaluate_string(
    prediction="Hello world",
    reference="Hello world!"
)

This code creates a StringEvaluator and evaluates a prediction against a reference string.

Execution Table

Step	Action	Input	Evaluation Logic	Output
1	Create StringEvaluator instance	None	Initialize evaluator	Evaluator ready
2	Call evaluate_string	prediction='Hello world', reference='Hello world!'	Compare strings with tolerance	Score calculated
3	Return evaluation result	Score calculated	Format result	{"score": 0.9, "feedback": "Close match"}
4	End	Evaluation complete	No further action	Evaluation finished

💡 Evaluation completes after returning the score and feedback.

Variable Tracker

Variable	Start	After Step 1	After Step 2	After Step 3	Final
eval	None	StringEvaluator instance	StringEvaluator instance	StringEvaluator instance	StringEvaluator instance
prediction	None	None	"Hello world"	"Hello world"	"Hello world"
reference	None	None	"Hello world!"	"Hello world!"	"Hello world!"
result	None	None	None	{"score": 0.9, "feedback": "Close match"}	{"score": 0.9, "feedback": "Close match"}

Key Moments - 2 Insights

Why does the evaluator return a score less than 1 even though the prediction looks very similar to the reference?

Because the evaluation logic compares strings exactly or with some tolerance, the missing exclamation mark causes a slight difference, resulting in a score less than 1 as shown in step 2 and 3 of the execution_table.

What happens if the prediction or reference is None or empty?

The evaluator will still run but may return a low or zero score because the comparison logic expects strings. This is implied in the input column of step 2 where valid strings are required.

Visual Quiz - 3 Questions

Look at the execution_table, what is the output after step 3?

A {"score": 0.9, "feedback": "Close match"}

B Evaluator ready

C Score calculated

D Evaluation finished

💡 Hint

Check the Output column in row with Step 3.

At which step does the evaluator compare the prediction and reference strings?

A Step 1

B Step 3

C Step 2

D Step 4

💡 Hint

Look at the Evaluation Logic column to find where comparison happens.

If the prediction was exactly the same as the reference, how would the score change in the execution_table?

A Score would be 0

B Score would be 1

C Score would be 0.5

D Score would be negative

💡 Hint

Consider the meaning of a perfect match in evaluation logic at step 2.

Concept Snapshot

LangSmith evaluators compare model outputs to references.
They run evaluation logic to produce scores or feedback.
Use evaluator methods like evaluate_string for text.
Outputs help measure model accuracy or quality.
Scores range from 0 (bad) to 1 (perfect match).

Full Transcript

LangSmith evaluators are tools that check how well a model's output matches a reference answer. The process starts by giving the evaluator the model's prediction and the correct reference. The evaluator then runs its logic to compare these two inputs. This comparison results in a score and sometimes feedback explaining the quality. For example, a StringEvaluator compares text strings and returns a score close to 1 if they are very similar. The evaluation ends by returning this score and feedback. This helps developers understand how accurate or good their model's outputs are.

Practice

1. What is the main purpose of LangSmith evaluators in LangChain?

easy

A. To check how good AI outputs are by comparing predictions to references

B. To train new AI models from scratch

C. To store large datasets for AI training

D. To create user interfaces for AI applications

Solution

1. Step 1: Understand the role of evaluators

   LangSmith evaluators are designed to assess AI outputs by comparing them with expected answers.

2. Step 2: Identify the correct purpose

   They do not train models, store data, or build interfaces but focus on evaluation.

3. Final Answer:

   To check how good AI outputs are by comparing predictions to references -> Option A

4. Quick Check:

   Evaluator purpose = Checking AI output quality [OK]

Hint: Evaluators compare AI answers to references to check quality [OK]

Common Mistakes:

• Confusing evaluators with training tools
• Thinking evaluators store data
• Assuming evaluators build UI

2. Which of the following is the correct way to call an evaluator's evaluate method in LangSmith?

easy

A. evaluate(evaluator, prediction, reference)

B. evaluator.evaluate(prediction, reference)

C. evaluator.run(reference, prediction)

D. evaluate(prediction, reference, evaluator)

Solution

1. Step 1: Recall method usage

   The evaluate method is called on the evaluator object with prediction and reference as arguments.

2. Step 2: Match correct syntax

   evaluator.evaluate(prediction, reference) matches this pattern exactly: evaluator.evaluate(prediction, reference).

3. Final Answer:

   evaluator.evaluate(prediction, reference) -> Option B

4. Quick Check:

   Method call = evaluator.evaluate(prediction, reference) [OK]

Hint: Call evaluate on evaluator with prediction and reference [OK]

Common Mistakes:

• Swapping argument order
• Calling evaluate as a standalone function
• Using wrong method name like run

3. Given the code snippet:

evaluator = SomeEvaluator()
prediction = "The sky is blue."
reference = "The sky is clear and blue."
result = evaluator.evaluate(prediction, reference)
print(result)

What is the expected behavior of print(result)?

medium

A. It prints the reference string unchanged

B. It prints the prediction string unchanged

C. It prints a score or feedback comparing prediction to reference

D. It raises a syntax error because evaluate needs more arguments

Solution

1. Step 1: Understand evaluate output

   The evaluate method returns a score or feedback about how close the prediction matches the reference.

2. Step 2: Analyze print statement

   Printing result shows this evaluation output, not the original strings or errors.

3. Final Answer:

   It prints a score or feedback comparing prediction to reference -> Option C

4. Quick Check:

   Evaluate returns score/feedback [OK]

Hint: Evaluate returns comparison result, not original text [OK]

Common Mistakes:

• Expecting evaluate to return input strings
• Thinking evaluate raises error without extra args
• Confusing prediction and reference outputs

4. What is the error in this code snippet?

evaluator = SomeEvaluator()
result = evaluator.evaluate(reference, prediction)
print(result)

Assuming evaluate expects (prediction, reference) order.

medium

A. Arguments are reversed; prediction should come before reference

B. Missing import statement for SomeEvaluator

C. evaluate method does not exist on evaluator

D. print statement syntax is incorrect

Solution

1. Step 1: Check argument order

   The evaluate method expects prediction first, then reference, but code reverses them.

2. Step 2: Confirm other parts are correct

   Assuming SomeEvaluator is imported and evaluate exists, the main issue is argument order.

3. Final Answer:

   Arguments are reversed; prediction should come before reference -> Option A

4. Quick Check:

   Correct argument order = prediction, reference [OK]

Hint: Remember evaluate(prediction, reference) argument order [OK]

Common Mistakes:

• Swapping prediction and reference arguments
• Assuming missing imports cause this error
• Thinking print syntax is wrong

5. You want to compare multiple AI model outputs to a single reference answer using LangSmith evaluators. Which approach correctly applies evaluators to get scores for each prediction?

hard

A. Combine all predictions into one string and evaluate against reference once

B. Call evaluator.evaluate once with a list of predictions and one reference

C. Use evaluator.evaluate(reference, prediction) inside a loop over references

D. Loop over predictions, call evaluator.evaluate(prediction, reference) for each, collect results

Solution

1. Step 1: Understand evaluator usage for multiple inputs

   Evaluators typically compare one prediction to one reference at a time.

2. Step 2: Apply evaluator in a loop

   Looping over each prediction and calling evaluate separately gives individual scores.

3. Step 3: Eliminate incorrect options

   Passing lists or combining strings is not standard; argument order matters.

4. Final Answer:

   Loop over predictions, call evaluator.evaluate(prediction, reference) for each, collect results -> Option D

5. Quick Check:

   Evaluate each prediction separately in a loop [OK]

Hint: Evaluate predictions one by one in a loop against reference [OK]

Common Mistakes:

• Passing lists instead of single strings
• Mixing argument order
• Combining predictions into one string