{% if eval_results.usage %}
{% from '_cache_kpis.html' import cache_kpis %}
{% set usage = eval_results.usage %}
{% set lanes = [('Answer generation', 'generation', 'retrieved_reader'), ('Gold-evidence diagnostic', 'oracle', 'oracle_reader'), ('Judge', 'judge', 'judge')] %}
{% set cache = namespace(prompt=0, cached=0, writes=0, reported=0, hits=0, tracked=false) %}
{% for label, prefix, lane in lanes %}
{% set cache.prompt = cache.prompt + (usage.get(prefix ~ '_prompt_tokens') or 0) %}
{% set cache.cached = cache.cached + (usage.get(prefix ~ '_cached_tokens') or 0) %}
{% set cache.writes = cache.writes + (usage.get(prefix ~ '_cache_write_tokens') or 0) %}
{% set cache.reported = cache.reported + (usage.get(prefix ~ '_cache_reported_calls') or 0) %}
{% set cache.hits = cache.hits + (usage.get(prefix ~ '_cache_hit_calls') or 0) %}
{% if (prefix ~ '_cache_reported_calls') in usage %}{% set cache.tracked = true %}{% endif %}
{% endfor %}
{% if cache.prompt %}
<section class="card eval-prompt-cache" id="eval-prompt-cache" aria-label="Provider prompt cache" style="padding:20px;margin:20px 0">
<h3 style="margin-top:0">Provider prompt cache</h3>
{{ cache_kpis((100 * cache.cached / cache.prompt)|round(1) if cache.tracked or cache.cached else none, cache.hits, cache.reported, cache.cached, cache.writes if cache.tracked else none) }}
<p class="usage-notice">Counts are provider-reported across answer, diagnostic and judge calls. Each case sends different evidence, so reuse comes mainly from shared instructions, and providers cache only prefixes above a model-specific minimum length. {% if not cache.tracked %}This run predates per-call cache tracking; the hit rate and writes were not recorded.{% endif %}</p>
</section>
{% endif %}
<details class="card" style="padding:20px;margin:20px 0"><summary>Token usage &amp; cost details</summary>
<p>Usage is reported by the provider. Missing prices or usage mean the cost is unknown. <a href="/evalground/compare">Compare models &amp; rerankers</a> for per-call measurements.</p>
<div style="overflow-x:auto"><table class="info-table"><thead><tr><th>Purpose</th><th>Input tokens</th><th>Cached input</th><th>Cache writes</th><th>Cache hit rate</th><th>Output tokens</th><th>API cost</th></tr></thead><tbody>
{% for label, prefix, lane in lanes %}
{% set reported = usage.get(prefix ~ '_cache_reported_calls') %}
<tr><td>{{ label }}</td><td>{{ usage.get(prefix ~ '_prompt_tokens', '—') }}</td><td>{{ usage.get(prefix ~ '_cached_tokens', '—') }}</td><td>{{ usage.get(prefix ~ '_cache_write_tokens', '—') }}</td><td>{% if reported %}{{ '%.1f%%'|format(100 * usage.get(prefix ~ '_cache_hit_calls', 0) / reported) }} <small>({{ usage.get(prefix ~ '_cache_hit_calls', 0) }}/{{ reported }})</small>{% else %}—{% endif %}</td><td>{{ usage.get(prefix ~ '_completion_tokens', '—') }}</td><td>{% set cost = usage.get('lane_cost_usd', {}).get(lane) %}{% if cost is not none %}${{ '%.6f'|format(cost) }}{% else %}Unknown{% endif %}</td></tr>
{% endfor %}
</tbody></table></div>
<details><summary>Raw usage &amp; pricing metadata</summary><pre style="white-space:pre-wrap;overflow-wrap:anywhere">{{ usage|tojson(indent=2) }}</pre></details>
</details>
{% endif %}