CachingConfig

CachingConfig controls Anthropic prompt caching across direct Anthropic, Azure Anthropic, and compatible Claude models on AWS Bedrock. Appam maps the same high-level setting onto the provider-specific caching mechanism underneath.

Struct Definition

pub struct CachingConfig {
    pub enabled: bool,
    pub ttl: CacheTTL,
}

pub enum CacheTTL {
    FiveMinutes,  // "5m" - default, refreshed on use
    OneHour,      // "1h" - 2x write cost
}

Fields

enabled

Whether prompt caching is active.

ttl

Cache time-to-live duration:

• CacheTTL::FiveMinutes -- 5-minute cache (default). Refreshed on each use, so active conversations maintain the cache automatically.
• CacheTTL::OneHour -- 1-hour cache. Higher write cost (2x base input token price) but longer retention.

Default

impl Default for CachingConfig {
    fn default() -> Self {
        Self {
            enabled: true,
            ttl: CacheTTL::FiveMinutes,
        }
    }
}

Provider Support

• Anthropic -- Supported. Appam maps CachingConfig to Anthropic's top-level cache_control field.
• Azure Anthropic -- Supported with the same request semantics as direct Anthropic.
• Bedrock -- Supported for compatible Claude models. Appam injects block-level cache_control checkpoints into supported Anthropic fields because Bedrock uses explicit cache checkpoints rather than Anthropic's top-level helper.

Cache Behavior

• Prefix Matching: Automatic prefix lookup (up to ~20 blocks)
• Direct Anthropic / Azure Anthropic: Anthropic applies the top-level marker to the last cacheable block in the request
• Bedrock: Appam injects block-level checkpoints into the end of the supported system, messages, and tools sections present in the request
• Refresh: 5-minute caches refresh on each hit, effectively staying alive for the duration of an active conversation

Pricing

Cache usage is tracked in AggregatedUsage:

• cache_creation_input_tokens -- Tokens used to create cache entries
• cache_read_input_tokens -- Tokens read from cache (where cost savings come from)

Builder Usage

use appam::prelude::*;
use appam::llm::anthropic::CachingConfig;

let agent = AgentBuilder::new("my-agent")
    .model("claude-sonnet-4-5")
    .caching(CachingConfig::default())
    .system_prompt("You are a helpful assistant with cached context.")
    .build()?;

TOML Configuration

[anthropic.caching]
enabled = true
ttl = "5m"

Source

Defined in src/llm/anthropic/config.rs.