Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
686 changes: 686 additions & 0 deletions src/proxy/guardrails.rs

Large diffs are not rendered by default.

80 changes: 78 additions & 2 deletions src/proxy/handlers/chat_completions/mod.rs
Original file line number Diff line number Diff line change
Expand Up @@ -10,16 +10,22 @@ use span_attributes::{
};
pub use types::ChatCompletionError;

use super::FormatHandlerAdapter;
use crate::{
gateway::{
error::GatewayError,
formats::OpenAIChatFormat,
traits::ProviderCapabilities,
traits::{ChatFormat, ProviderCapabilities},
types::{
common::Usage,
openai::{ChatCompletionChunk, ChatCompletionRequest, ChatCompletionResponse},
},
},
proxy::handlers::format_handler::FormatHandlerAdapter,
proxy::guardrails::{
input_guardrail_payload_from_chat_messages, input_payload_from_check_payload,
input_payload_to_chat_messages, output_guardrail_payload_from_chat_messages,
output_payload_from_check_payload, output_payload_to_chat_messages,
},
};

pub(crate) struct ChatCompletionsAdapter;
Expand Down Expand Up @@ -86,7 +92,77 @@ impl FormatHandlerAdapter for ChatCompletionsAdapter {
collector.output_message_span_properties()
}

fn guardrail_input_payload(
_lifecycle_state: &Self::LifecycleState,
request: &Self::Request,
) -> Result<Option<crate::guardrail::traits::GuardrailCheckPayload>, Self::Error> {
let (hub_request, _) = OpenAIChatFormat::to_hub(request)?;
let payload = input_guardrail_payload_from_chat_messages(&hub_request.messages)
.map(crate::guardrail::traits::GuardrailCheckPayload::Input)
.map_err(bridge_error)?;
Ok(Some(payload))
}

fn apply_input_guardrail_rewrite(
_lifecycle_state: &mut Self::LifecycleState,
request: &mut Self::Request,
rewrite: crate::guardrail::traits::GuardrailCheckPayload,
) -> Result<(), Self::Error> {
request.messages = input_payload_to_chat_messages(
&input_payload_from_check_payload(rewrite).map_err(bridge_error)?,
)
.map_err(bridge_error)?;
Ok(())
}

fn guardrail_output_payload(
_lifecycle_state: &Self::LifecycleState,
response: &Self::Response,
) -> Result<Option<crate::guardrail::traits::GuardrailCheckPayload>, Self::Error> {
let messages = response
.choices
.iter()
.map(|choice| choice.message.clone())
.collect::<Vec<_>>();
let payload = output_guardrail_payload_from_chat_messages(&messages)
.map(crate::guardrail::traits::GuardrailCheckPayload::Output)
.map_err(bridge_error)?;
Ok(Some(payload))
}

fn apply_output_guardrail_rewrite(
_lifecycle_state: &mut Self::LifecycleState,
response: &mut Self::Response,
rewrite: crate::guardrail::traits::GuardrailCheckPayload,
) -> Result<(), Self::Error> {
let messages = output_payload_to_chat_messages(
&output_payload_from_check_payload(rewrite).map_err(bridge_error)?,
)
.map_err(bridge_error)?;

if messages.len() != response.choices.len() {
return Err(bridge_error(GatewayError::Bridge(format!(
"chat completion output guardrail rewrite expected {} messages, got {}",
response.choices.len(),
messages.len()
))));
}

for (choice, message) in response.choices.iter_mut().zip(messages) {
choice.message = message;
}

Ok(())
}

fn end_of_stream_event(saw_item: bool) -> Option<SseEvent> {
saw_item.then(|| SseEvent::default().data("[DONE]"))
}
}

fn bridge_error<E>(error: E) -> ChatCompletionError
where
E: std::fmt::Display,
{
ChatCompletionError::GatewayError(GatewayError::Bridge(error.to_string()))
}
116 changes: 116 additions & 0 deletions src/proxy/handlers/format_handler.rs
Original file line number Diff line number Diff line change
Expand Up @@ -20,6 +20,7 @@ use tokio::{

use crate::{
config::entities::{Model, ResourceEntry},
guardrail::traits::{GuardrailCheckPayload, GuardrailOutcome},
gateway::{
error::GatewayError,
traits::{ChatFormat, ProviderCapabilities},
Expand All @@ -30,6 +31,7 @@ use crate::{
},
proxy::{
AppState,
guardrails::{ConfiguredGuardrailRuntime, resolve_model_guardrails},
hooks::{
self, RequestContext, authorization::AuthorizationError, rate_limit::RateLimitError,
},
Expand Down Expand Up @@ -85,6 +87,36 @@ pub(crate) trait FormatHandlerAdapter: Send + Sync + 'static {

fn output_message_span_properties(collector: &Self::Collector) -> Vec<(String, String)>;

fn guardrail_input_payload(
_lifecycle_state: &Self::LifecycleState,
_request: &Self::Request,
) -> Result<Option<GuardrailCheckPayload>, Self::Error> {
Ok(None)
}

fn apply_input_guardrail_rewrite(
_lifecycle_state: &mut Self::LifecycleState,
_request: &mut Self::Request,
_rewrite: GuardrailCheckPayload,
) -> Result<(), Self::Error> {
Ok(())
}

fn guardrail_output_payload(
_lifecycle_state: &Self::LifecycleState,
_response: &Self::Response,
) -> Result<Option<GuardrailCheckPayload>, Self::Error> {
Ok(None)
}

fn apply_output_guardrail_rewrite(
_lifecycle_state: &mut Self::LifecycleState,
_response: &mut Self::Response,
_rewrite: GuardrailCheckPayload,
) -> Result<(), Self::Error> {
Ok(())
}

async fn prepare_lifecycle(
_state: &AppState,
_request_ctx: &mut RequestContext,
Expand Down Expand Up @@ -183,11 +215,19 @@ where
let provider = model.provider(resources.as_ref()).ok_or_else(|| {
GatewayError::Internal(format!("provider {} not found", model.provider_id))
})?;
let configured_guardrails = resolve_model_guardrails(&model, resources.as_ref())?;
let provider_instance = create_provider_instance(gateway.as_ref(), &provider)?;
let provider_base_url = provider_instance.effective_base_url().ok();
let mut lifecycle_state =
A::prepare_lifecycle(&state, &mut request_ctx, &mut request_data).await?;

apply_input_guardrails::<A>(
&configured_guardrails,
&mut lifecycle_state,
&mut request_data,
)
.await?;

let span = Span::enter_with_local_parent(A::span_name());
apply_span_properties(
&span,
Expand All @@ -214,6 +254,12 @@ where
mut response,
usage,
})) => {
apply_output_guardrails::<A>(
&configured_guardrails,
&mut lifecycle_state,
&mut response,
)
.await?;
A::handle_complete_response(
&state,
&mut request_ctx,
Expand Down Expand Up @@ -247,6 +293,76 @@ where
}
}

async fn apply_input_guardrails<A>(
guardrails: &[Box<dyn ConfiguredGuardrailRuntime>],
lifecycle_state: &mut A::LifecycleState,
request: &mut AdapterRequest<A>,
) -> Result<(), A::Error>
where
A: FormatHandlerAdapter,
{
for guardrail in guardrails {
let Some(payload) = A::guardrail_input_payload(lifecycle_state, request)? else {
continue;
};
let Some(outcome) = guardrail.check(&payload).await? else {
continue;
};

match outcome {
GuardrailOutcome::Allow => {}
GuardrailOutcome::Rewrite(rewrite) => {
A::apply_input_guardrail_rewrite(lifecycle_state, request, rewrite)?;
}
GuardrailOutcome::Block { reason } => {
return Err(GatewayError::Validation(format!(
"guardrail {} blocked input: {}",
guardrail.name(),
reason
))
.into());
}
}
}

Ok(())
}

async fn apply_output_guardrails<A>(
guardrails: &[Box<dyn ConfiguredGuardrailRuntime>],
lifecycle_state: &mut A::LifecycleState,
response: &mut AdapterResponse<A>,
) -> Result<(), A::Error>
where
A: FormatHandlerAdapter,
{
for guardrail in guardrails {
let Some(payload) = A::guardrail_output_payload(lifecycle_state, response)? else {
continue;
};
let Some(outcome) = guardrail.check(&payload).await? else {
continue;
};

match outcome {
GuardrailOutcome::Allow => {}
GuardrailOutcome::Rewrite(rewrite) => {
A::apply_output_guardrail_rewrite(lifecycle_state, response, rewrite)?;
}
GuardrailOutcome::Block { reason } => {
return Err(GatewayError::Validation(format!(
"guardrail {} blocked output: {}",
guardrail.name(),
reason
))
.into());
}
}
}

Ok(())
}

async fn handle_regular_response<A>(
response: AdapterResponse<A>,
usage: Usage,
Expand Down
Loading