Speakr is a personal, self-hosted web application designed for transcribing audio recordings
Python
Document prompt-cache reads in admin statistics and the summary/chat timestamp availability toggles in account settings
Murtaza Nasir786a2f1ff717
diff --git a/docs/admin-guide/statistics.md b/docs/admin-guide/statistics.mdindex 4944f8af..e000cbca 100644--- a/docs/admin-guide/statistics.md+++ b/docs/admin-guide/statistics.md@@ -48,6 +48,8 @@ The Token Usage section provides visibility into LLM API consumption across your- Yellow: Between 80-100% (warning zone)- Red: At or over 100% (blocked)+**Prompt Cache Reads**: The LLM card also reports prompt-cache activity for the current month (and today): **cache reads** are input tokens that the model served from a prefix cache instead of recomputing, and **writes** are tokens written into that cache. These figures only appear when the text-generation backend reports them in its API response. They are most useful alongside the `PREFIX_CACHE_OPTIMIZED_PROMPTS` option on a self-hosted backend with prefix caching, where the summary call reuses the transcript that the title call already prefilled. On hosted APIs that do not report cached tokens, the figure stays at zero even if caching is happening upstream. See [Prefix-Cache-Friendly Prompts](model-configuration.md#prefix-cache-friendly-prompts-prefix_cache_optimized_prompts) for how to enable this and what flags your server needs.+**Cost Tracking**: When using OpenRouter or other providers that return cost information, the statistics include estimated costs based on actual API responses. For embeddings, the cost is calculated from the configured per-million-token price for the active embedding provider. This helps with budgeting and identifying high-cost operations.Use token statistics to identify heavy users, validate budget allocations, and forecast API costs. If certain users consistently hit their limits, you may need to increase their budgets or investigate their usage patterns. See [Token Budget Management](user-management.md#token-budget-management) for setting individual user limits.diff --git a/docs/user-guide/settings.md b/docs/user-guide/settings.mdindex fb0e9c30..2f28e358 100644--- a/docs/user-guide/settings.md+++ b/docs/user-guide/settings.md@@ -83,6 +83,15 @@ These defaults apply to all your recordings unless overridden by tag defaults, f!!! tip "When to use hotwords vs initial prompt"Use **hotwords** for specific terms the model tends to misspell - proper nouns, brand names, technical acronyms. Use **initial prompt** for broader context that helps the model understand the domain and make better overall word choices.+### Timestamp Availability for Summary and Chat++At the top of the Prompt Options you can control whether the AI receives per-line timestamps with the transcript. There are two independent toggles, one for **summarization** and one for **chat**, so you can enable the feature only where you need it:++- **Summary** - when on, the summarizer sees the time of each line and can reference moments in the recording (for example, "the budget was approved around 00:14:30"). Useful for long meetings where locating a decision matters.+- **Chat** - when on, the chat assistant can answer time-aware questions such as "what was discussed near the start" or "around the 20-minute mark".++Each toggle has a format dropdown. **Default** uses a compact `[hh:mm:ss] Speaker: text` layout; you can also select any of your saved transcript templates to control exactly how each line is rendered. Both toggles are off by default, which keeps the prompt shorter and avoids spending tokens on timestamps when you do not need them.+### Writing Effective PromptsCraft prompts based on how you actually use summaries. Do you extract action items for project management? Look for decisions that affect strategy? Track technical details for documentation? Your prompt should request exactly what you need for these next steps.