Skip to content

Switchyard lets LLM applications route traffic across models and providers while preserving native OpenAI and Anthropic API compatibility - enabling flexible model selection, benchmarking, and cost/performance optimization.

Rust

fix(libsy): put judge system prompt in instructions, not messages (#312)

Lin Jiaf4aa2889f584
diff --git a/crates/libsy/src/algorithms/llm_class.rs b/crates/libsy/src/algorithms/llm_class.rs
index 6367374dd..b22a0166d 100644
--- a/crates/libsy/src/algorithms/llm_class.rs
+++ b/crates/libsy/src/algorithms/llm_class.rs
@@ -919,7 +919,8 @@ mod tests {
use super::*;
use switchyard_protocol::{
- LlmClientError, LlmRequest, Metadata, completion_text, text_request, text_response,
+ ContentBlock, InstructionBlock, LlmClientError, LlmRequest, Metadata, completion_text,
+ text_request, text_response,
};
use crate::algorithms::util::llm_judge::Judge;
@@ -1003,9 +1004,17 @@ mod tests {
self.judge_system_prompts.lock().extend(
request
.llm_request
- .messages
+ .instructions
.first()
- .and_then(|message| message.text_content("\n")),
+ .and_then(|instruction| {
+ instruction.content.iter().find_map(|b| {
+ if let ContentBlock::Text { text } = b {
+ Some(text.clone())
+ } else {
+ None
+ }
+ })
+ }),
);
r#"{"crux":"bounded task","primary_rule":"SUP-1","capability_boundary":"supported","p_solve":0.9}"#.to_string()
} else {
@@ -1473,7 +1482,17 @@ mod tests {
let judge_request = judge.build_request(&State::default(), &request);
assert_eq!(judge_request.llm_request.model, request.llm_request.model);
- assert_eq!(judge_request.llm_request.messages.len(), 3);
+ assert_eq!(judge_request.llm_request.instructions.len(), 1);
+ assert_eq!(judge_request.llm_request.instructions[0].role, Role::System);
+ assert_eq!(
+ judge_request.llm_request.instructions[0].content,
+ InstructionBlock {
+ role: Role::System,
+ content: Message::text(Role::System, judge.contract().system_prompt()).content,
+ }
+ .content,
+ );
+ assert_eq!(judge_request.llm_request.messages.len(), 2);
let contents = judge_request
.llm_request
.messages
diff --git a/crates/libsy/src/algorithms/util/escalation.rs b/crates/libsy/src/algorithms/util/escalation.rs
index 49cd53085..b8387f625 100644
--- a/crates/libsy/src/algorithms/util/escalation.rs
+++ b/crates/libsy/src/algorithms/util/escalation.rs
@@ -375,12 +375,13 @@ mod tests {
.push(Message::text(Role::Assistant, "this turn's reply"));
let built = judge.build_request(&State::default(), &judged);
- // Two messages: the rubric as system, the condensed trajectory as user.
- assert_eq!(built.llm_request.messages.len(), 2);
- assert_eq!(built.llm_request.messages[0].role, Role::System);
- assert_eq!(built.llm_request.messages[1].role, Role::User);
+ // Rubric in instructions, condensed trajectory as the sole user message.
+ assert_eq!(built.llm_request.instructions.len(), 1);
+ assert_eq!(built.llm_request.instructions[0].role, Role::System);
+ assert_eq!(built.llm_request.messages.len(), 1);
+ assert_eq!(built.llm_request.messages[0].role, Role::User);
assert!(
- built.llm_request.messages[1]
+ built.llm_request.messages[0]
.text_content("")
.is_some_and(|text| text.contains("Conversation turn 4"))
);
diff --git a/crates/libsy/src/algorithms/util/llm_judge.rs b/crates/libsy/src/algorithms/util/llm_judge.rs
index 90430e5a2..3898c44fb 100644
--- a/crates/libsy/src/algorithms/util/llm_judge.rs
+++ b/crates/libsy/src/algorithms/util/llm_judge.rs
@@ -14,7 +14,7 @@ use async_trait::async_trait;
use serde::de::DeserializeOwned;
use serde_json::Value;
use switchyard_protocol::{
- AggLlmResponse, LlmRequest, Message, OutputParams, Role, completion_text,
+ AggLlmResponse, InstructionBlock, LlmRequest, Message, OutputParams, Role, completion_text,
};
use super::classifier_contract::ClassifierContract;
@@ -144,14 +144,15 @@ where
type Verdict = D::Verdict;
fn build_request(&self, state: &State, request: &Request) -> Request {
- let mut messages = self.input.build_messages(state, request);
- messages.insert(
- 0,
- Message::text(Role::System, self.contract.system_prompt().to_string()),
- );
+ let messages = self.input.build_messages(state, request);
Request {
llm_request: LlmRequest {
model: request.llm_request.model.clone(),
+ instructions: vec![InstructionBlock {
+ role: Role::System,
+ content: Message::text(Role::System, self.contract.system_prompt().to_string())
+ .content,
+ }],
messages,
output: OutputParams {
max_output_tokens: Some(self.runtime.max_output_tokens),
diff --git a/tests/test_libsy_minimal_bindings.py b/tests/test_libsy_minimal_bindings.py
index b8f0a25c7..b3182cd1c 100644
--- a/tests/test_libsy_minimal_bindings.py
+++ b/tests/test_libsy_minimal_bindings.py
@@ -100,7 +100,7 @@ async def test_classifier_config_accepts_a_prompt_override() -> None:
_, response = await algorithm.run(request_body())
- prompt = judge.calls[0]["messages"][0]["content"][0]["text"]
+ prompt = judge.calls[0]["instructions"][0]["content"][0]["text"]
assert prompt == "Custom capability rubric."
assert judge.calls[0]["output"]["response_format"]["json_schema"]["schema"][
"properties"