{% extends "base.html" %} {% block title %}Command palette searches{% endblock %} {% block content %}

Command palette searches

{{ total_searches }} searches logged · {{ total_bounces }} with no results. Add common bounce terms as synonyms on an existing command palette page, or create a new page for them.

{% if llm_calls %}
Natural-language assist

{{ llm_calls }} model call{{ llm_calls|pluralize }} · {{ llm_total_tokens }} tokens ({{ llm_prompt_tokens }} prompt / {{ llm_completion_tokens }} completion){% if llm_failures %} · {{ llm_failures }} failed{% endif %}.

{% comment %} The prompt total on its own reads as far more expensive than it is. The system prompt is the same few thousand tokens on every call, so the provider serves nearly all of it from cache at a fraction of the normal input rate. Rounds-per-query is the number worth watching: it multiplies everything above, and a figure well over 1 means the model is being asked again and again. {% endcomment %}

{{ llm_cached_prompt_tokens }} of those prompt tokens ({{ llm_cached_percent }}%) were served from the provider's cache and are billed at a fraction of the normal input rate — only {{ llm_uncached_prompt_tokens }} were charged at full price.

{{ llm_rounds_per_query }} model calls per query on average. Each round re-sends the whole system prompt, so this is the multiplier on everything above.

{% if llm_by_action %}

{% for row in llm_by_action %}{{ row.action }} ×{{ row.count }}{% if not forloop.last %} · {% endif %}{% endfor %}

{% endif %}
{% endif %} {% if llm_reported_queries %}
Reported by the people who hit them
{% comment %} Every other list on this page is inferred: we watched what happened and worked out that it went badly. This one is a person reading a failure and deciding it was worth pressing a button about, which makes it the shortest queue here and the first one worth reading. {% endcomment %}

{{ llm_reported }} command(s) reported with “that didn't work”. Nothing was emailed — this is the queue.

{% for row in llm_reported_queries %} {% endfor %}
QueryWhat it triedCount
{{ row.query }} {{ row.action|default:row.response_kind }} {{ row.count }}
{% endif %} {% if llm_cancelled_queries %}
Commands people stopped
{% comment %} These are worse than the failures below, and much harder to find any other way. The assistant was confident, the server was happy, the countdown started -- and the person watching it hit Cancel. Nothing ran, so there is no error and no history entry: this table is the only record that the query was understood as the wrong thing. {% endcomment %}

Cancelled during the countdown: {{ llm_cancelled }} ({{ llm_cancelled_percent }}% of confirmable actions). Nothing was written. A phrase repeated here was matched to the wrong action, or to the right action on the wrong auction.

{% for row in llm_cancelled_queries %} {% endfor %}
QueryWas going to runCount
{{ row.query }} {{ row.action }} {{ row.count }}
{% endif %} {% if llm_gave_up %}
Queries the assistant couldn't answer
{% comment %} This is the feature backlog: a phrase repeated here is somebody asking for a skill that doesn't exist. {% endcomment %}

Each of these fell through to search results or a guess. A phrase that shows up repeatedly is a request for an action that isn't in auctions/palette_actions.py yet, or a page missing keywords in auctions/palette_routes.py.

{% for row in llm_gave_up %} {% endfor %}
QueryOutcomeReportedCount
{{ row.query }} {{ row.response_kind }} {% if row.reports %}{{ row.reports }}{% endif %} {{ row.count }}
{% endif %}
Top searches
{% for row in top_searches %} {% empty %} {% endfor %}
SearchCountClicks
{{ row.search }}{{ row.count }}{{ row.clicks }}
No searches yet.
Top bounces (no results — candidates for new shortcuts/synonyms)
{% for row in top_bounces %} {% empty %} {% endfor %}
SearchCount
{{ row.search }}{{ row.count }}
No bounces yet.
{% endblock %}