Eval harness — recall@k, faithfulness, latency, cost

import json, time
from anthropic import Anthropic

client = Anthropic()

def load_golden(path="evals/golden.jsonl"):
    with open(path) as f:
        return [json.loads(line) for line in f]

def eval_one(case, search_fn):
    t0 = time.perf_counter()
    chunks = search_fn(case["question"])
    answer = answer_fn(case["question"], chunks)
    duration_s = time.perf_counter() - t0

    retrieved_ids = {c["id"] for c in chunks[:5]}
    expected_ids = set(case["expected_chunk_ids"])
    recall_at_5 = len(retrieved_ids & expected_ids) / max(len(expected_ids), 1)

    # LLM-judge for answer accuracy
    judge = client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=10,
        messages=[{"role": "user", "content": f"Does this answer contain the expected fact? Yes or No.\n\nAnswer: {answer['text']}\n\nExpected: {case['expected_answer']}"}],
    )
    accuracy = 1 if "yes" in judge.content[0].text.lower() else 0

    return {
        "case_id": case["id"],
        "recall_at_5": recall_at_5,
        "accuracy": accuracy,
        "latency_s": duration_s,
        "cost_usd": estimate_cost(...),
    }

def run_eval():
    cases = load_golden()
    results = [eval_one(c, hybrid_then_rerank) for c in cases]
    return {
        "recall_at_5": sum(r["recall_at_5"] for r in results) / len(results),
        "accuracy": sum(r["accuracy"] for r in results) / len(results),
        "p99_latency_s": sorted([r["latency_s"] for r in results])[int(0.99 * len(results))],
        "avg_cost_usd": sum(r["cost_usd"] for r in results) / len(results),
    }