{# ============================ HEADER ============================ #}

{% if p.run.org_label %}{{ p.run.org_label }}{% else %}Bedrock model invocation logs{% endif %}

source {{ p.source.type }} {% if p.source.window.first_seen %} window {{ p.source.window.first_seen }} → {{ p.source.window.last_seen }} {% endif %} generated {{ p.run.generated_at[:19] | replace('T', ' ') }} semantr {{ p.generator.version }} schema {{ p.schema_version }}

total spend

use cases

conversations

invocations

per conversation upper bound

tokens

Spend composition

{% for seg in composition %} {% endfor %}

{% for seg in composition %}

{{ seg.name }} {{ seg.share | pct }}

{% endfor %}

{% if sparkline %}

Daily spend

{{ sparkline | safe }} {% endif %}

Work profile

{% for cell in quadrants %}

{{ cell.use_cases }} use case{{ 's' if cell.use_cases != 1 }} · {{ cell.sessions | num }} conv

{% endfor %}

Rows: software work vs everything else. Columns: templated or scheduled work vs one-off and exploratory. Cadence is measured from prompt repetition and how many days the work appears on, not inferred by a model.

{# ============================ TABS ============================ #} {# ---------------------------- USE CASES ---------------------------- #}

Discovered use cases

Conversations grouped by similarity, each group named from its own contents. No fixed taxonomy and no other bucket. Select a row for detail.

{% if p.use_cases %}

{% for tag in ['coding', 'non-coding', 'repeatable', 'experimental'] %} {% endfor %}

{{ p.use_cases | length }} of {{ p.use_cases | length }}

{% for u in p.use_cases %} {% endfor %}

	Use case	Tags	Conversations	Spend	Share
{{ loop.index }}	▶ {{ u.name }} {{ u.description }}	{% for tag in u.tags %}{{ tag }}{% endfor %} {% if u.confidence < 0.5 %}low confidence{% endif %} {% if not u.named %}unnamed{% endif %}	{{ u.sessions \| num }}	{{ u.cost_usd \| money }}	{{ u.cost_share \| pct(1) }}
Per conversation {{ u.cost_per_session \| money }} Invocations {{ u.invocations \| num }} Distinct prompts {{ u.distinct_prompts \| num }} Active days {{ u.active_days \| num }} Tokens in / out {{ u.input_tokens \| num }} / {{ u.output_tokens \| num }} First → last seen {{ u.first_seen or '—' }} → {{ u.last_seen or '—' }} Model {% if u.models %}{{ u.models[0].model \| shortmodel }}{% else %}—{% endif %} Application {% if u.apps %}{{ u.apps[0].app }}{% else %}—{% endif %} Category {{ u.category }} Confidence {{ u.confidence \| pct }} {% if u.tools %} Tools called {% for t in u.tools %}{{ t.tool }}{% if not loop.last %} · {% endif %}{% endfor %} {% endif %} {% if u.recommendations %} Cost improvements {% for r in u.recommendations %} {{ r.title }} {% if r.annual_saving_usd %}— up to {{ r.annual_saving_usd \| money }}/yr{% endif %} {{ r.detail }} {% endfor %} {% endif %} {% if u.examples %} Sample prompts {% for e in u.examples %} {{ e }} {% endfor %} {% endif %}

{% else %}

No use cases were discovered — the source contained no readable invocations.

{% endif %}

{# ---------------------------- DETAILS ---------------------------- #}

Details

What was actually invoked — the models, tools, applications and IAM principals present in the logs, with their counts.

Models — {{ p.spend.by_model | length }}

{% for b in p.spend.by_model %} {% endfor %}

Model	Conversations	Invocations	Input tokens	Output tokens	Spend
{{ b.key \| shortmodel }} {{ b.key }}	{{ b.sessions \| num }}	{{ b.invocations \| num }}	{{ b.input_tokens \| num }}	{{ b.output_tokens \| num }}	{{ b.cost_usd \| money }}

Tools — {{ p.spend.by_tool | length }}

{% if p.spend.by_tool %}

{% for b in p.spend.by_tool %} {% endfor %}

Tool	Conversations	Invocations	Spend
{{ b.key }}	{{ b.sessions \| num }}	{{ b.invocations \| num }}	{{ b.cost_usd \| money }}

{% else %}

No tool definitions appeared in these logs.

{% endif %}

Applications — {{ p.spend.by_app | length }}

{% for b in p.spend.by_app %} {% endfor %}

Execution role	Conversations	Invocations	Spend
{{ b.key }}	{{ b.sessions \| num }}	{{ b.invocations \| num }}	{{ b.cost_usd \| money }}

IAM principals — {{ p.spend.by_principal | length }}

{% for b in p.spend.by_principal %} {% endfor %}

Principal ARN	Conversations	Spend
{{ b.key }}	{{ b.sessions \| num }}	{{ b.cost_usd \| money }}

Account ids are replaced with a stable hash. {% if p.contains_prompt_text %} This file contains verbatim prompt text — the sample prompts above are quoted from your logs. Check them before forwarding it outside the company. The Share button strips them; a shared link never carries prompt text. {% else %} No prompt text is included — this file is safe to forward. {% endif %}

{% if p.source.cost_coverage < 1 %}

Pricing was unavailable for {{ (1 - p.source.cost_coverage) | pct(1) }} of invocations ({{ p.source.unpriced_models.keys() | join(', ') }}). Those are costed at zero, so every spend figure in this report is a floor.

{% endif %}

{# ---------------------------- RISK ---------------------------- #}

Risk flags

Retrospective detection over retained prompts. Counts and labels only — matched values are never written into this file.

{% if not p.risk %}

Nothing flagged.

{% else %}

{% for f in p.risk %} {% endfor %}

Type	Finding	Top application	Conversations	Occurrences
{{ f.kind }}	{{ f.label }}	{% if f.top_apps %}{{ f.top_apps[0].app }}{% if f.top_apps \| length > 1 %} +{{ f.top_apps \| length - 1 }}{% endif %}{% else %}—{% endif %}	{{ f.sessions \| num }}	{{ f.occurrences \| num }}

{% endif %}

{# ---------------------------- CONVERSATIONS ---------------------------- #}

Costliest conversations

Ranked by whole conversation rather than by request. A twenty-turn session is one unit of work; tools that count requests never show it that way.

{% for s in p.costliest_sessions %} {% endfor %}

Spend	Invocations	Turns	Model	Application	Opening prompt
{{ s.cost_usd \| money }}	{{ s.invocations \| num }}	{{ s.turns \| num }}	{{ s.model \| shortmodel }}	{{ s.app }}	{{ s.first_turn \| default('—', true) \| truncate(100) }}

{# ---------------------------- RUN ---------------------------- #}

Run

{{ p.source.ingest.records_read | num }} invocation records were joined into {{ p.totals.sessions | num }} conversations, reduced to {{ p.totals.distinct_snippets | num }} distinct prompts, and grouped into {{ p.totals.clusters }} clusters. Each cluster was named once — that is {{ p.run.inference.complete_calls | num }} model call{{ 's' if p.run.inference.complete_calls != 1 }}, regardless of corpus size.

How to read the conversation count. Bedrock logs one invocation at a time and records no conversation id, so conversations are rebuilt by matching each request against the one whose output it contains. The join is exact — no similarity, no guessing — but three things split one real conversation into several: a retry, a context compaction, and an assistant branch the user abandoned. So {{ p.totals.sessions | num }} is an upper bound, and against a synthetic corpus with known ground truth it runs about 10% high. Everything derived from it — cost per conversation, conversations per use case — is a lower bound in the same proportion. Spend is unaffected: it is summed per invocation and never divided by this number.

Analysis cost	{{ p.run.inference.cost_usd \| money }}
Naming calls	{{ p.run.inference.complete_calls \| num }}
Cached input tokens	{{ p.run.inference.cached_input_tokens \| num }}
Records read	{{ p.source.ingest.records_read \| num }}
Records skipped	{{ p.source.ingest.records_skipped \| num }}
Own calls excluded	{{ p.source.ingest.excluded_own_calls \| num }}
Cost coverage	{{ p.source.cost_coverage \| pct(1) }}
Snippet strategy	{{ p.run.config.snippet_strategy }}
Granularity	{{ p.run.config.granularity }}
Run time	{{ p.run.duration_seconds }}s
Run id	{{ p.run.id }}

{% if p.source.ingest.skip_reasons %}

Skipped: {% for reason, n in p.source.ingest.skip_reasons.items() %}{{ reason }} ({{ n | num }}){% if not loop.last %}, {% endif %}{% endfor %}.

{% endif %} {% for note in p.run.inference.notes %}

{% endfor %}

{% if p.run.org_label %}{{ p.run.org_label }}{% else %}Bedrock model invocation logs{% endif %}

Spend composition

Daily spend

Work profile

Discovered use cases

Tools called

Cost improvements

Sample prompts

Details

Models — {{ p.spend.by_model | length }}

Tools — {{ p.spend.by_tool | length }}

Applications — {{ p.spend.by_app | length }}

IAM principals — {{ p.spend.by_principal | length }}

Risk flags

Costliest conversations

Run