LangChain Eval Harness (Python)

Overview

A team swapped gpt-4o for claude-sonnet-4-6 to save money and a week later CS noticed answer quality dropped on 15% of refund tickets — the regression was invisible in code review and invisible in CI because no golden set existed.

Fix: a versioned golden set, a stacked eval pipeline (LangSmith + ragas + deepeval + custom trajectory), and a PR-blocking regression gate with paired Wilcoxon significance. The tooling exists; the patterns for wiring it into a statistically honest loop are scattered across five doc sites.

Build a 100-example JSONL golden set, wire LangSmith evaluate() with a custom correctness evaluator, add a ragas quartet (faithfulness, answer relevance, context precision/recall) for RAG, add deepeval LLM-as-judge with N=3 judge quorum, score LangGraph trajectories on coverage/precision/ order, and gate PRs on a 2% aggregate drop or 5% per-example drop. Pin: langchain-core 1.0.x, langgraph 1.0.x, langsmith>=0.2, ragas>=0.2, deepeval>=2.0. Pain-catalog anchors: P01, P11, P12, P22, P33.

Prerequisites

• Python 3.10+
• langchain-core >= 1.0, < 2.0, langgraph >= 1.0, < 2.0 for the system under eval
• pip install langsmith>=0.2 ragas>=0.2 deepeval>=2.0 scipy
• LangSmith account + LANGSMITH_API_KEY (free tier is sufficient for dataset versioning)
• Provider API keys for the judge LLM: OPENAI_API_KEY and/or ANTHROPIC_API_KEY

Instructions

Step 1 — Build a versioned golden set

Format: JSONL, one example per line, with a dataset_version tag. Minimum 20 examples to start; grow to 100 for PR gating, 200+ for absolute-metric claims.

python
# evals/golden_set/v2026.04.jsonl
{"id": "gs-0001", "input": "Refund policy for SKU ABC-42?", "expected": "30 days with receipt", "contexts": ["policy_v3.md"], "tags": ["refund"], "difficulty": "easy", "dataset_version": "2026.04"}
{"id": "gs-0002", "input": "Return policy for opened software?", "expected": "No, opened software is final sale", "contexts": ["policy_v3.md#returns"], "tags": ["refund"], "difficulty": "medium", "dataset_version": "2026.04"}

Sample from real traffic (redacted), not imagination. Stratify by tag and difficulty (aim for 30% hard). Two annotators per example, disagreements reconciled — reconciliation rate under 90% means your task definition is ambiguous. Treat the file as immutable within a version; bump the version to refresh. See Golden Set Curation for sourcing strategy, annotation tool options, and the refresh cadence.

Step 2 — Wire LangSmith evaluate() with a custom evaluator

python
from langsmith import Client
from langsmith.evaluation import evaluate, EvaluationResult
from langchain_anthropic import ChatAnthropic

client = Client()
DATASET_VERSION = "2026.04"

# One-time: upload golden set as a versioned dataset
def upload_golden_set(jsonl_path, dataset_name):
    examples = [json.loads(line) for line in open(jsonl_path)]
    client.create_dataset(dataset_name)
    client.create_examples(
        inputs=[{"input": e["input"]} for e in examples],
        outputs=[{"expected": e["expected"]} for e in examples],
        metadata=[{"id": e["id"], "tags": e["tags"]} for e in examples],
        dataset_name=dataset_name,
    )

chain = ChatAnthropic(model="claude-sonnet-4-6", temperature=0, timeout=30)

def target(inputs):
    return {"answer": chain.invoke(inputs["input"]).content}

def correctness(outputs, reference_outputs):
    """Deterministic exact-match floor — baseline, not ceiling."""
    match = outputs["answer"].strip().lower() == reference_outputs["expected"].strip().lower()
    return EvaluationResult(key="exact_match", score=float(match))

results = evaluate(
    target,
    data=f"golden-set-v{DATASET_VERSION}",
    evaluators=[correctness],
    experiment_prefix="refund-bot-v3",
    max_concurrency=10,   # Avoid 429s on judge LLM (P22)
)

Free-form outputs need semantic scoring (ragas, deepeval, or LLM-as-judge — Step 4).

Step 3 — Add ragas metrics for RAG pipelines

For a RAG chain returning {answer, contexts}, ragas scores four standard dimensions. The default judge is gpt-4o-mini; override to pin model + cost:

python
from ragas import evaluate as ragas_evaluate
from ragas.metrics import faithfulness, answer_relevancy, context_precision, context_recall
from langchain_openai import ChatOpenAI
from langchain_openai import OpenAIEmbeddings
from datasets import Dataset

judge = ChatOpenAI(model="gpt-4o-mini", temperature=0)
embed = OpenAIEmbeddings(model="text-embedding-3-small")

# Prepare rows — ragas wants HuggingFace Dataset shape
rows = []
for ex in golden_examples:
    result = rag_chain.invoke({"question": ex["input"]})
    rows.append({
        "question": ex["input"],
        "answer": result["answer"],
        "contexts": [d.page_content for d in result["source_documents"]],
        "ground_truth": ex["expected"],
    })

ragas_results = ragas_evaluate(
    Dataset.from_list(rows),
    metrics=[faithfulness, answer_relevancy, context_precision, context_recall],
    llm=judge,
    embeddings=embed,
)
# ragas_results is a dict of per-metric means; call .to_pandas() for per-row

Do not use ragas on non-RAG chains — context_precision against an empty context list returns 0 and looks like a regression. See Framework Comparison for when each tool fits.

Step 4 — Add deepeval LLM-as-judge for free-form outputs

deepeval is pytest-shaped — each example is an LLMTestCase asserting against metrics. Run N=3 judge invocations per example and take the median to tame LLM-as-judge variance (±5-15% across runs; single-run scores are not CI-ready):

python
import statistics
from deepeval.metrics import GEval
from deepeval.test_case import LLMTestCase, LLMTestCaseParams

def eval_with_quorum(test_case, metric, n=3):
    scores = []
    for _ in range(n):
        metric.measure(test_case)
        scores.append(metric.score)
    return statistics.median(scores), statistics.stdev(scores) if n > 1 else 0.0

correctness = GEval(
    name="Correctness",
    criteria="Does the actual output match the expected output in meaning?",
    evaluation_params=[LLMTestCaseParams.INPUT, LLMTestCaseParams.ACTUAL_OUTPUT, LLMTestCaseParams.EXPECTED_OUTPUT],
    model="gpt-4o-mini",
)

for ex in golden_examples:
    result = chain.invoke({"input": ex["input"]})
    case = LLMTestCase(input=ex["input"], actual_output=result, expected_output=ex["expected"])
    median, sd = eval_with_quorum(case, correctness, n=3)
    if sd > 0.2:  # judge disagreeing with itself — flag, don't gate
        flag_for_review(ex["id"], median, sd)

Step 5 — LangGraph agent trajectory eval

For agents, final-answer correctness misses the process. Score the tool-call sequence on three axes — coverage (did required tools run?), precision (were extra tools used?), and order (Kendall's tau on shared tools):

python
from langchain_core.messages import AIMessage

def extract_trajectory(final_state: dict) -> list[dict]:
    return [
        {"tool": tc["name"], "args": tc["args"]}
        for msg in final_state["messages"] if isinstance(msg, AIMessage)
        for tc in (msg.tool_calls or [])
    ]

def trajectory_score(expected: list[str], actual: list[str]) -> dict:
    e_set, a_set = set(expected), set(actual)
    coverage = len(e_set & a_set) / len(e_set) if e_set else 1.0
    precision = len(e_set & a_set) / len(a_set) if a_set else 0.0
    shared = [t for t in actual if t in e_set]
    order = _kendall_tau(expected, shared) if len(shared) >= 2 else 1.0
    return {"coverage": coverage, "precision": precision, "order": order}

# Composite: 0.5 * coverage + 0.3 * precision + 0.2 * order

Set temperature=0 for the agent during eval — temperature > 0 produces different trajectories across runs (P11) and makes paired comparison statistically invalid. See Agent Trajectory Eval for args-level matching, efficiency/safety scoring, and the LLM-as-judge fallback for non-deterministic trajectories.

Step 6 — Gate PRs on regression

A PR touching prompts, chain code, or model config runs the eval suite on PR branch and main, then blocks merge on any of: aggregate mean drop > 2.0%, any single-example drop > 5.0%, or paired Wilcoxon signed-rank p < 0.05 with negative mean delta.

python
from scipy.stats import wilcoxon

def paired_regression_check(baseline, candidate, alpha=0.05):
    """Wilcoxon — right test when metric distribution is non-normal (most LLM metrics)."""
    n = len(baseline)
    if n < 50:
        return {"verdict": "too_small_n", "n": n}
    diffs = [c - b for b, c in zip(baseline, candidate)]
    _, p = wilcoxon(diffs, alternative="less")
    return {"n": n, "mean_delta": sum(diffs) / n, "p_value": float(p),
            "regression": p < alpha and sum(diffs) < 0}

At n=100 and α=0.05 this detects a ~3-5% true regression at ~80% power. See CI Integration for the GitHub Actions workflow, PR-comment delta table, bootstrap CI, and spend/rate-limit safety rails.

Output

• JSONL golden set at evals/golden_set/v2026.04.jsonl with an immutable version tag
• LangSmith dataset uploaded and versioned; experiment runs linked to traces
• Ragas scores (faithfulness, answer relevance, context precision/recall) on RAG chains
• Deepeval LLMTestCase assertions in pytest, with median-of-3 judge quorum
• LangGraph trajectory scores (coverage, precision, order) with composite summary
• GitHub Actions workflow gating PRs on 2% aggregate / 5% per-example / Wilcoxon p < 0.05
• PR-comment delta table posted on every eval run

Framework selection at a glance

| Use case | LangSmith | ragas | deepeval | Custom | |---|---|---|---|---| | RAG metrics (faithfulness, context recall) | — | Primary | Fallback | — | | Pytest-style assertion in CI | Secondary | — | Primary | — | | Trace capture + dataset versioning | Primary | Complementary | Complementary | — | | Agent trajectory (tool-call sequence) | Secondary (traces) | — | — | Primary | | Exact match / JSON schema / structured output | — | — | — | Primary | | Free-form paraphrase scoring | Via custom evaluator | — | Primary (G-Eval) | — |

Most real pipelines stack two or three. The anti-pattern is running all four on every example — you pay $10-30 per run for signal you are not using. See Framework Comparison for the full decision tree and dependency weight comparison.

Error Handling

| Error / Failure mode | Cause | Fix | |---|---|---| | TimeoutError on eval runs > 20 min | Long agent trajectories on slow models; 100 examples × 30s each exceeds default GH Actions job timeout | Cap max_concurrency=10, use asyncio.gather with asyncio.Semaphore, split eval into sharded jobs | | Judge disagreement (stdev > 0.2 on 0,1] scale across N=3 runs) | LLM-as-judge variance on ambiguous examples | Flag example for manual review; do not use that row's score for gating | | ValidationError: missing 'contexts' in ragas | Chain does not return retrieved docs | Modify chain to surface source_documents, or switch to non-RAG evaluator | | Wilcoxon p-value is NaN | All paired diffs are 0 (identical outputs) | Expected when the PR did not change behavior — no regression, skip the stat test | | LangSmith 429 rate limit during upload | > 50 examples/sec to create_examples | Batch with client.create_examples(..., batch_size=20) and sleep between batches | | Spend overrun ($50+ per run) | Judge calls scaling with N_examples × N_metrics × N_judge_runs | Use gpt-4o-mini not gpt-4o for judge; cache per (dataset_version, chain_version) | | AttributeError: 'list' has no attribute 'lower' in custom evaluator | Claude AIMessage.content is list[dict] not str (P02 — see langchain-model-inference) | Use msg.text() or iterate content blocks | | Trajectory comparison drifts week-over-week on unchanged agent | temperature > 0 non-determinism (P11) | Set temperature=0 for all eval runs; pin seed where supported |

Examples

Setting up eval for a new RAG chain

Start with 20 production-sampled golden examples, wire up ragas_evaluate with four metrics, record scores to evals/baselines/ as the reference, and promote to LangSmith dataset versioning once two engineers annotate in parallel. See Golden Set Curation.

Diagnosing regression after a model swap

Run the main-branch chain on the golden set, then swap the model and rerun. Diff per-example scores sorted by delta — the top-10 regressions usually cluster by tag (long contexts, one-shot lookups). Report paired Wilcoxon and per-tag breakdown before deciding to ship. See CI Integration.

Evaluating a LangGraph tool-calling agent

Record expected tool-call sequences for 50 tasks, capture actual trajectories via extract_trajectory, and score on coverage/precision/order. Composite drops indicate a policy change — diff sequences to find the drift. See Agent Trajectory Eval.

Resources

• LangSmith evaluation tutorial
• LangSmith evaluate() reference
• ragas metrics overview
• deepeval metrics reference
• scipy Wilcoxon signed-rank
• G-Eval variance paper
• Pack pain catalog: docs/pain-catalog.md (entries P01, P11, P12, P22, P33)