langsmith-evaluator（朗史密斯评估器）

LANGSMITH_API_KEY=lsv2_pt_your_api_key_here          # Required
LANGSMITH_WORKSPACE_ID=your-workspace-id              # Optional: for org-scoped keys
OPENAI_API_KEY=your_openai_key                        # For LLM as Judge

Python Dependencies

pip install langsmith langchain-openai python-dotenv

JavaScript Dependencies

npm install langsmith commander chalk cli-table3 dotenv openai

<crucial_requirement>

Golden Rule: Inspect Before You Implement

CRITICAL: Before writing ANY evaluator or extraction logic, you MUST:

1. Run your agent on sample inputs and capture the actual output
2. Inspect the output - print it, query LangSmith traces, understand the exact structure
3. Only then write code that processes that output

Output structures vary significantly by framework, agent type, and configuration. Never assume the shape - always verify first. Query LangSmith traces to when outputs don't contain needed data to understand how to extract from execution. </crucial_requirement>

<evaluator_format>

Offline vs Online Evaluators

Offline Evaluators (attached to datasets):

• Function signature: (run, example) - receives both run outputs and dataset example
• Use case: Comparing agent outputs to expected values in a dataset
• Upload with: --dataset "Dataset Name"

Online Evaluators (attached to projects):

• Function signature: (run) - receives only run outputs, NO example parameter
• Use case: Real-time quality checks on production runs (no reference data)
• Upload with: --project "Project Name"

CRITICAL - Return Format:

• Return {"score": value, "comment": "..."} - the metric key is auto-derived from the function name
• Each evaluator returns ONE metric only. For multiple metrics, create multiple evaluator functions.
• Do NOT return {"metric_name": value} or lists of metrics - this will error.

CRITICAL - Local vs Uploaded (Python only):

• Local evaluate(): run is a RunTree object → use run.outputs
• Uploaded to LangSmith: run is a dict → use run["outputs"]
• Handle both: run.outputs if hasattr(run, "outputs") else run.get("outputs", {})
• TypeScript always uses attribute access: run.outputs?.field </evaluator_format>

<evaluator_types>

• LLM as Judge - Uses an LLM to grade outputs. Best for subjective quality (accuracy, helpfulness, relevance).
• Custom Code - Deterministic logic. Best for objective checks (exact match, trajectory validation, format compliance). </evaluator_types>

<llm_judge>

LLM as Judge Evaluators

NOTE: LLM-as-Judge upload is currently not supported by our script only supports code evaluators. For evaluations against a dataset, STRONGLY PREFER defining local evaluators to use with evaluate(evaluators=[...]).

class Grade(TypedDict): reasoning: Annotated[str,..., "Explain your reasoning"] is_accurate: Annotated[bool,..., "True if response is accurate"]

judge = ChatOpenAI(model="gpt-4o-mini", temperature=0).with_structured_output(Grade, method="json_schema", strict=True)

async def accuracy_evaluator(run, example): run_outputs = run.outputs if hasattr(run, "outputs") else run.get("outputs", {}) or {} example_outputs = example.outputs if hasattr(example, "outputs") else example.get("outputs", {}) or {} grade = await judge.ainvoke([{"role": "user", "content": f"Expected: {example_outputs}\nActual: {run_outputs}\nIs this accurate?"}]) return {"score": 1 if grade["is_accurate"] else 0, "comment": grade["reasoning"]}

</python>
</llm_judge>

<code_evaluators>
## Custom Code Evaluators

**Before writing an evaluator:**
1. Inspect your dataset to understand expected field names (see Golden Rule above)
2. Test your run function and verify its output structure matches the dataset schema
3. Query LangSmith traces to debug any mismatches

<python>

def trajectory_evaluator(run, example): run_outputs = run.outputs if hasattr(run, "outputs") else run.get("outputs", {}) or {} example_outputs = example.outputs if hasattr(example, "outputs") else example.get("outputs", {}) or {} # IMPORTANT: Replace these placeholders with your actual field names # 1. Query your LangSmith trace to see what fields exist in run outputs # 2. Check your dataset schema for expected field names # Note: Trajectory data may not appear in default output - verify against trace! actual = run_outputs.get("YOUR_TRAJECTORY_FIELD", []) expected = example_outputs.get("YOUR_EXPECTED_FIELD", []) return {"score": 1 if actual == expected else 0, "comment": f"Expected {expected}, got {actual}"}

<run_functions>

## Defining Run Functions

Run functions execute your agent and return outputs for evaluation.

**CRITICAL - Test Your Run Function First:** Before writing evaluators, you MUST test your run function and inspect the actual output structure. Output shapes vary by framework, agent type, and configuration.

**Debugging workflow:**

1. Run your agent once on sample input
2. Query the trace to see the execution structure
3. Print the raw output and verify against trace to output contains the right data
4. Adjust the run function as needed
5. Verify your output matches your dataset schema

**Try your hardest to match your run function output to your dataset schema.** This makes evaluators simple and reusable. If matching isn't possible, your evaluator must know how to extract and compare the right fields from each side.

def run_agent(inputs: dict) -> dict: result = your_agent.run(inputs) # ALWAYS inspect output shape first - run this, check the print, query traces print(f"DEBUG - type: {type(result)}, keys: {result.keys() if hasattr(result, 'keys') else 'N/A'}") print(f"DEBUG - value: {result}") return {"output": result} # Adjust to match your dataset schema

### Capturing Trajectories

For trajectory evaluation, your run function must capture tool calls during execution.

**CRITICAL:** Run output formats vary significantly by framework and agent type. You MUST inspect before implementing:

**LangGraph agents (LangChain OSS):** Use `stream_mode="debug"` with `subgraphs=True` to capture nested subagent tool calls.

import uuid

def run_agent_with_trajectory(agent, inputs: dict) -> dict: config = {"configurable": {"thread_id": f"eval-{uuid.uuid4()}"}} trajectory = [] final_result = None

for chunk in agent.stream(inputs, config=config, stream_mode="debug", subgraphs=True): # STEP 1: Print chunks to understand the structure print(f"DEBUG chunk: {chunk}")

# STEP 2: Write extraction based on YOUR observed structure # ... your extraction logic here ...

# IMPORTANT: After running, query the LangSmith trace to verify # your trajectory data is complete. Default output may be missing # tool calls that appear in the trace. return {"output": final_result, "trajectory": trajectory}

**Custom / Non-LangChain Agents:**

1. **Inspect output first** - Run your agent and inspect the result structure. Trajectory data may already be included in the output (e.g., `result.tool_calls`, `result.steps`, etc.)
2. **Callbacks/Hooks** - If your framework supports execution callbacks, register a hook that records tool names on each invocation
3. **Parse execution logs** - As a last resort, extract tool names from structured logs or trace data

The key is to capture the tool name at execution time, not at definition time. </run_functions>

**IMPORTANT - Auto-Run Behavior:** Evaluators uploaded to a dataset **automatically run** when you run experiments on that dataset. You do NOT need to pass them to `evaluate()` - just run your agent against the dataset and the uploaded evaluators execute automatically.

**IMPORTANT - Local vs Uploaded:** Uploaded evaluators have very limited package access for security reasons! DO NOT upload evaluators that unless they only need to rely on standard Python / Javascript functionality, such as built-in packages. For dataset (offline) evaluators, prefer running locally with `evaluate(evaluators=[...])` first. This gives you full package access.

**IMPORTANT - Code vs Structured Evaluators:**

- **Code evaluators** (what our script uploads): Run in a limited environment without external packages. Use for deterministic logic (exact match, trajectory validation).
- **Structured evaluators** (LLM-as-Judge): Configured via LangSmith UI, use a specific payload format with model/prompt/schema. Our script does not support this format yet.

**IMPORTANT - Choose the right target:**

- `--dataset`: Offline evaluator with `(run, example)` signature - for comparing to expected values
- `--project`: Online evaluator with `(run)` signature - for real-time quality checks

You must specify one. Global evaluators are not supported.

List all evaluators

python upload_evaluators.py list

Upload offline evaluator (attached to dataset)

python upload_evaluators.py upload my_evaluators.py \ --name "Trajectory Match" --function trajectory_evaluator \ --dataset "My Dataset" --replace

Upload online evaluator (attached to project)

python upload_evaluators.py upload my_evaluators.py \ --name "Quality Check" --function quality_check \ --project "Production Agent" --replace

Delete

python upload_evaluators.py delete "Trajectory Match"

List all evaluators

npx tsx upload_evaluators.ts list

Upload offline evaluator (attached to dataset)

npx tsx upload_evaluators.ts upload my_evaluators.js \ --name "Trajectory Match" --function trajectoryEvaluator \ --dataset "My Dataset" --replace

Upload online evaluator (attached to project)

npx tsx upload_evaluators.ts upload my_evaluators.js \ --name "Quality Check" --function qualityCheck \ --project "Production Agent" --replace

Delete

npx tsx upload_evaluators.ts delete "Trajectory Match"

**IMPORTANT - Safety Prompts:**

- The script prompts for confirmation before destructive operations
- **NEVER use `--yes` flag unless the user explicitly requests it**

<best_practices>

1. **Use structured output for LLM judges** - More reliable than parsing free-text
2. **Match evaluator to dataset type**
  - Final Response → LLM as Judge for quality
  - Trajectory → Custom Code for sequence
3. **Use async for LLM judges** - Enables parallel evaluation
4. **Test evaluators independently** - Validate on known good/bad examples first
5. **Choose the right language**
  - Python: Use for Python agents, langchain integrations
  - JavaScript: Use for TypeScript/Node.js agents </best_practices>

<running_evaluations>

## Running Evaluations

**Uploaded evaluators** auto-run when you run experiments - no code needed. **Local evaluators** are passed directly for development/testing.

from langsmith import evaluate

Uploaded evaluators run automatically

results = evaluate(run_agent, data="My Dataset", experiment_prefix="eval-v1")

Or pass local evaluators for testing

results = evaluate(run_agent, data="My Dataset", evaluators=[my_evaluator], experiment_prefix="eval-v1")

</running_evaluations>

**Output doesn't match what you expect:** Query the LangSmith trace. It shows exact inputs/outputs at each step - compare what you find to what you're trying to extract.

**One metric per evaluator:** Return `{"score": value, "comment": "..."}`. For multiple metrics, create separate functions.

**Field name mismatch:** Your run function output must match dataset schema exactly. Inspect dataset first with `client.read_example(example_id)`.

**RunTree vs dict (Python):** Local `evaluate()` passes `RunTree`, uploaded evaluators receive `dict`. Handle both:

run_outputs = run.outputs if hasattr(run, "outputs") else run.get("outputs", {}) or {}