langgraph-testing-evaluation（语言图测试评估）

LangGraph Testing & Evaluation

Practical workflows for validating agent quality with:

• Unit/integration tests
• Trajectory evaluation
• LangSmith dataset evaluations
• A/B-style comparisons between versions

Use this file for high-level flow. Load references/* for detailed implementation.

Start Here

Choose the smallest approach that answers your question:

Recommended order:

1. Unit tests
2. Integration/trajectory checks
3. Dataset evaluation in LangSmith
4. A/B comparison before deployment

Quick Commands

Run from repo root.

Generate test scaffolding

# Python (preferred)
uv run skills/langgraph-testing-evaluation/scripts/generate_test_cases.py my_agent:graph --output tests/ --framework pytest

# JavaScript/TypeScript
node skills/langgraph-testing-evaluation/scripts/generate_test_cases.js ./my-agent.ts:graph --output tests/ --framework vitest

Run trajectory evaluation

# Python: LLM-as-judge
uv run skills/langgraph-testing-evaluation/scripts/run_trajectory_eval.py my_agent:run_agent my_dataset --method llm-judge --model openai:o3-mini

# Python: trajectory match
uv run skills/langgraph-testing-evaluation/scripts/run_trajectory_eval.py my_agent:run_agent dataset.json --method match --trajectory-match-mode strict --reference-trajectory reference.json

# JavaScript/TypeScript
node skills/langgraph-testing-evaluation/scripts/run_trajectory_eval.js ./agent.ts:runAgent my_dataset --method llm-judge --model openai:o3-mini --max-concurrency 4

Run LangSmith dataset evaluation

# Python
uv run skills/langgraph-testing-evaluation/scripts/evaluate_with_langsmith.py my_agent:run_agent my_dataset --evaluators accuracy,latency --max-concurrency 4

# Python (do not upload experiment results)
uv run skills/langgraph-testing-evaluation/scripts/evaluate_with_langsmith.py my_agent:run_agent my_dataset --evaluators accuracy --no-upload

# JavaScript/TypeScript
node skills/langgraph-testing-evaluation/scripts/evaluate_with_langsmith.js ./agent.ts:runAgent my_dataset --evaluators accuracy,latency --max-concurrency 4

Compare two agent versions

# Python
uv run skills/langgraph-testing-evaluation/scripts/compare_agents.py my_agent:v1 my_agent:v2 dataset.json --output comparison_report.json

# JavaScript/TypeScript
node skills/langgraph-testing-evaluation/scripts/compare_agents.js ./v1.ts:run ./v2.ts:run dataset.json --output comparison_report.json

# JavaScript/TypeScript (force local dataset file only)
node skills/langgraph-testing-evaluation/scripts/compare_agents.js ./v1.ts:run ./v2.ts:run dataset.json --no-langsmith

Create mock response configs

# Python
uv run skills/langgraph-testing-evaluation/scripts/mock_llm_responses.py create --type sequence --output mock_config.json

# JavaScript/TypeScript
node skills/langgraph-testing-evaluation/scripts/mock_llm_responses.js create --type sequence --output mock_config.json

Core Workflow

1. Define test scope.

• Unit: deterministic logic in one node/function.
• Integration: node interactions and routing.
• End-to-end: complete response quality on realistic inputs.

1. Start from deterministic checks.

• Mock LLM/tool IO for speed and repeatability.
• Keep real-model tests as a smaller, explicit suite.

1. Build/curate dataset examples.

• Use stable inputs and expected outputs.
• Keep schema simple: inputs and outputs objects (optional metadata).
• Compatibility note: scripts also accept singular keys (input, output) for legacy datasets.

1. Run evaluation with explicit gates.

• Use evaluator keys that map to deployment decisions.
• Set thresholds in CI for regression prevention.

1. Compare versions before rollout.

• Run same dataset on both versions.
• Check both quality and latency.

1. Diagnose failures from traces/experiments.

• Inspect low-scoring examples.
• Split failures by pattern (routing, tool usage, hallucination, latency spikes).

Current References (Load On Demand)

references/unit-testing-patterns.md

Load when:

• You need node-level and routing test patterns.
• You need pytest/vitest/Jest integration patterns.
• You need robust mocking and flaky-test reduction.

references/trajectory-evaluation.md

Load when:

• You need trajectory match evaluation (strict, unordered, subset, superset).
• You need LLM-as-judge trajectory scoring.
• You need LangSmith experiment comparison for trajectory results.

references/langsmith-evaluation.md

Load when:

• You need dataset creation/management in LangSmith.
• You need evaluator signatures and experiment runs in Python/TS.
• You need CI-friendly workflows with quality thresholds.

references/ab-testing.md

Load when:

• You need offline A/B comparison methodology.
• You need significance testing and interpretation.
• You need production traffic split strategy and guardrails.

Assets

assets/templates/test_template.py

• Runnable Python pytest template aligned with current LangGraph testing patterns.
• Includes:

- Compiled-graph invocation with thread_id - Single-node testing via compiled_graph.nodes[...] - Integration-test placeholder

assets/datasets/sample_dataset.json

• Deterministic seed dataset for LangSmith ingestion.
• Uses examples: [{inputs, outputs, metadata}] format.

assets/examples/README.md

• Documentation-only index for current asset usage.
• Notes where runnable assets live today.

Script Interface Summary

scripts/generate_test_cases.py / .js

Use for fast test scaffolding.

Inputs:

• Graph module path

- Python: my_module:graph or my_module.graph - JS/TS: ./file.ts:graph

Outputs:

• Framework-specific starter tests in target directory.

scripts/run_trajectory_eval.py / .js

Use for trajectory scoring with either:

• --method match
• --method llm-judge

Supports:

• Local dataset files (.json)
• LangSmith dataset names
• Optional reference trajectory file with --reference-trajectory
• Match modes: strict, unordered, subset, superset

Local-only mode:

• --no-langsmith in both Python and JavaScript scripts (requires local JSON dataset file)

scripts/evaluate_with_langsmith.py / .js

Use for dataset-based evaluation runs and experiment tracking.

Supports:

• Existing dataset by name
• Dataset creation from JSON examples file
• Multiple evaluators (--evaluators accuracy,latency,...)
• Concurrency control (--max-concurrency)

Python-only:

• --no-upload to run without uploading experiment results

scripts/compare_agents.py / .js

Use for offline version comparisons:

• Shared dataset input
• Success/latency summaries
• JSON report output for CI artifacts
• Local JSON datasets or LangSmith datasets (JS supports --no-langsmith to disable remote loading)

scripts/mock_llm_responses.py / .js

Use for deterministic test doubles:

• single
• sequence
• conditional

Decision Rules

If behavior is deterministic and local:

• Use unit tests first.

If behavior depends on tool sequence/routing:

• Add trajectory evaluation.

If behavior depends on realistic distribution quality:

• Run LangSmith dataset evaluation.

If approving a replacement model/prompt/graph:

• Run A/B comparison and check both quality and latency.

Common Failure Patterns

Flaky tests

• Cause: real-model nondeterminism in unit scope.
• Fix: mock LLM/tool calls for unit tests; reserve real-model tests for separate integration marks.

High trajectory variance

• Cause: overly strict matching for workflows with equivalent paths.
• Fix: switch match mode (unordered, subset, or superset) where appropriate.

Regressions hidden by averages

• Cause: only aggregate score monitored.
• Fix: inspect per-example failures and segment by category metadata.

Latency regressions with same quality

• Cause: no explicit latency gate.
• Fix: include latency evaluator and CI threshold.

Minimal Best Practices

1. Keep fast deterministic tests as the largest share.
2. Version datasets and keep them stable.
3. Track both correctness and latency.
4. Add explicit go/no-go thresholds in CI.
5. Compare candidate vs baseline before production rollout.
6. Investigate failures with trace-level evidence, not only aggregate scores.