bedrock: Add system-prompt cache anchor on caching-capable models (#56474)

The Bedrock Converse API supports placing `CachePoint` blocks inside the
`system` field, but we were sending the system prompt as a single
`SystemContentBlock::Text`, which leaves the system tokens dependent on
whatever message-level breakpoint happens to fall within the 20-block
lookback window.

This widens `bedrock::Request.system` from `Option<String>` to
`Vec<BedrockSystemContentBlock>` and has `into_bedrock` emit
`[Text(system), CachePoint(Default)]` whenever the model supports prompt
caching. The system prompt now anchors its own cache prefix, on top of
the existing tool-list anchor and per-message breakpoint, so a stable
system prompt keeps producing cache hits even when earlier conversation
turns change.

Bedrock does not support automatic caching or the 1-hour TTL, so the
default 5-minute ephemeral cache is the only option for this provider.

Release Notes:

- Improved Bedrock prompt cache utilization by anchoring the system
prompt as its own cache prefix

crates/bedrock/src/bedrock.rs

@@ -84,10 +84,8 @@ pub async fn stream_completion(
 
     response = response.inference_config(inference_config);
 
-    if let Some(system) = request.system {
-        if !system.is_empty() {
-            response = response.system(BedrockSystemContentBlock::Text(system));
-        }
+    for system_block in request.system {
+        response = response.system(system_block);
     }
 
     if let Some(guardrail_id) = &request.guardrail_identifier {
@@ -207,7 +205,11 @@ pub struct Request {
     pub messages: Vec<BedrockMessage>,
     pub tools: Option<BedrockToolConfig>,
     pub thinking: Option<Thinking>,
-    pub system: Option<String>,
+    /// System content blocks in prefix order. Typically `[Text(...)]` or, when
+    /// the model supports prompt caching, `[Text(...), CachePoint(...)]` so the
+    /// system prompt anchors its own cache prefix independent of tools and
+    /// messages.
+    pub system: Vec<BedrockSystemContentBlock>,
     pub metadata: Option<Metadata>,
     pub stop_sequences: Vec<String>,
     pub temperature: Option<f32>,

crates/language_models/src/provider/bedrock.rs

@@ -7,6 +7,7 @@ use aws_config::stalled_stream_protection::StalledStreamProtectionConfig;
 use aws_config::{BehaviorVersion, Region};
 use aws_credential_types::{Credentials, Token};
 use aws_http_client::AwsHttpClient;
+use bedrock::BedrockSystemContentBlock;
 use bedrock::bedrock_client::Client as BedrockClient;
 use bedrock::bedrock_client::config::timeout::TimeoutConfig;
 use bedrock::bedrock_client::types::{
@@ -1104,11 +1105,24 @@ pub fn into_bedrock(
         )
     };
 
+    let mut system_blocks: Vec<BedrockSystemContentBlock> = Vec::new();
+    if !system_message.is_empty() {
+        system_blocks.push(BedrockSystemContentBlock::Text(system_message));
+        if supports_caching {
+            system_blocks.push(BedrockSystemContentBlock::CachePoint(
+                CachePointBlock::builder()
+                    .r#type(CachePointType::Default)
+                    .build()
+                    .context("failed to build system cache point block")?,
+            ));
+        }
+    }
+
     Ok(bedrock::Request {
         model,
         messages: new_messages,
         max_tokens: max_output_tokens,
-        system: Some(system_message),
+        system: system_blocks,
         tools: tool_config,
         thinking: if request.thinking_allowed {
             match thinking_mode {