apunkt/plano
1
0
Fork 0
mirror of https://github.com/katanemo/plano.git synced 2026-05-11 00:32:42 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

model routing: cost/latency ranking with ranked fallback list (#849)

Browse source
This commit is contained in:
Adil Hafeez 2026-03-30 13:46:52 -07:00 • committed by GitHub
parent 3a531ce22a
commit e5751d6b13
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
23 changed files with 1524 additions and 317 deletions
Download patch file Download diff file Expand all files Collapse all files

26
crates/brightstaff/src/handlers/llm/mod.rs
View file

@ -119,7 +119,7 @@ async fn llm_chat_inner(
temperature,
tool_names,
user_message_preview,
inline_routing_policy,
inline_routing_preferences,
client_api,
provider_id,
} = parsed;
@ -261,7 +261,7 @@ async fn llm_chat_inner(
&traceparent,
&request_path,
&request_id,
inline_routing_policy,
inline_routing_preferences,
)
.await
}
@ -323,7 +323,7 @@ struct PreparedRequest {
temperature: Option<f32>,
tool_names: Option<Vec<String>>,
user_message_preview: Option<String>,
inline_routing_policy: Option<Vec<common::configuration::ModelUsagePreference>>,
inline_routing_preferences: Option<Vec<common::configuration::TopLevelRoutingPreference>>,
client_api: Option<SupportedAPIsFromClient>,
provider_id: hermesllm::ProviderId,
}
@ -352,16 +352,14 @@ async fn parse_and_validate_request(
"request body received"
);
// Extract routing_policy from request body if present
let (chat_request_bytes, inline_routing_policy) =
crate::handlers::routing_service::extract_routing_policy(&raw_bytes, false).map_err(
|err| {
warn!(error = %err, "failed to parse request JSON");
let mut r = Response::new(full(format!("Failed to parse request: {}", err)));
*r.status_mut() = StatusCode::BAD_REQUEST;
r
},
)?;
// Extract routing_preferences from request body if present
let (chat_request_bytes, inline_routing_preferences) =
crate::handlers::routing_service::extract_routing_policy(&raw_bytes).map_err(|err| {
warn!(error = %err, "failed to parse request JSON");
let mut r = Response::new(full(format!("Failed to parse request: {}", err)));
*r.status_mut() = StatusCode::BAD_REQUEST;
r
})?;
let api_type = SupportedAPIsFromClient::from_endpoint(request_path).ok_or_else(|| {
warn!(path = %request_path, "unsupported endpoint");
@ -439,7 +437,7 @@ async fn parse_and_validate_request(
temperature,
tool_names,
user_message_preview,
inline_routing_policy,
inline_routing_preferences,
client_api,
provider_id,
})

36
crates/brightstaff/src/handlers/llm/model_selection.rs
View file

@ -1,6 +1,6 @@
use common::configuration::ModelUsagePreference;
use common::configuration::TopLevelRoutingPreference;
use hermesllm::clients::endpoints::SupportedUpstreamAPIs;
use hermesllm::{ProviderRequest, ProviderRequestType};
use hermesllm::ProviderRequestType;
use hyper::StatusCode;
use std::sync::Arc;
use tracing::{debug, info, warn};
@ -10,7 +10,10 @@ use crate::streaming::truncate_message;
use crate::tracing::routing;
pub struct RoutingResult {
/// Primary model to use (first in the ranked list).
pub model_name: String,
/// Full ranked list — use subsequent entries as fallbacks on 429/5xx.
pub models: Vec<String>,
pub route_name: Option<String>,
}
@ -39,11 +42,8 @@ pub async fn router_chat_get_upstream_model(
traceparent: &str,
request_path: &str,
request_id: &str,
inline_usage_preferences: Option<Vec<ModelUsagePreference>>,
inline_routing_preferences: Option<Vec<TopLevelRoutingPreference>>,
) -> Result<RoutingResult, RoutingError> {
// Clone metadata for routing before converting (which consumes client_request)
let routing_metadata = client_request.metadata().clone();
// Convert to ChatCompletionsRequest for routing (regardless of input type)
let chat_request = match ProviderRequestType::try_from((
client_request,
@ -78,22 +78,6 @@ pub async fn router_chat_get_upstream_model(
"router request"
);
// Use inline preferences if provided, otherwise fall back to metadata extraction
let usage_preferences: Option<Vec<ModelUsagePreference>> = if inline_usage_preferences.is_some()
{
inline_usage_preferences
} else {
let usage_preferences_str: Option<String> =
routing_metadata.as_ref().and_then(|metadata| {
metadata
.get("plano_preference_config")
.map(|value| value.to_string())
});
usage_preferences_str
.as_ref()
.and_then(|s| serde_yaml::from_str(s).ok())
};
// Prepare log message with latest message from chat request
let latest_message_for_log = chat_request
.messages
@ -107,7 +91,6 @@ pub async fn router_chat_get_upstream_model(
let latest_message_for_log = truncate_message(&latest_message_for_log, 50);
info!(
has_usage_preferences = usage_preferences.is_some(),
path = %request_path,
latest_message = %latest_message_for_log,
"processing router request"
@ -121,7 +104,7 @@ pub async fn router_chat_get_upstream_model(
.determine_route(
&chat_request.messages,
traceparent,
usage_preferences,
inline_routing_preferences,
request_id,
)
.await;
@ -132,10 +115,12 @@ pub async fn router_chat_get_upstream_model(
match routing_result {
Ok(route) => match route {
Some((route_name, model_name)) => {
Some((route_name, ranked_models)) => {
let model_name = ranked_models.first().cloned().unwrap_or_default();
current_span.record("route.selected_model", model_name.as_str());
Ok(RoutingResult {
model_name,
models: ranked_models,
route_name: Some(route_name),
})
}
@ -147,6 +132,7 @@ pub async fn router_chat_get_upstream_model(
Ok(RoutingResult {
model_name: "none".to_string(),
models: vec!["none".to_string()],
route_name: None,
})
}

152
crates/brightstaff/src/handlers/routing_service.rs
View file

@ -1,5 +1,5 @@
use bytes::Bytes;
use common::configuration::{ModelUsagePreference, SpanAttributes};
use common::configuration::{SpanAttributes, TopLevelRoutingPreference};
use common::consts::REQUEST_ID_HEADER;
use common::errors::BrightStaffError;
use hermesllm::clients::SupportedAPIsFromClient;
@ -15,56 +15,42 @@ use crate::handlers::llm::model_selection::router_chat_get_upstream_model;
use crate::router::llm::RouterService;
use crate::tracing::{collect_custom_trace_attributes, operation_component, set_service_name};
const ROUTING_POLICY_SIZE_WARNING_BYTES: usize = 5120;
/// Extracts `routing_policy` from a JSON body, returning the cleaned body bytes
/// and parsed preferences. The `routing_policy` field is removed from the JSON
/// before re-serializing so downstream parsers don't see the non-standard field.
///
/// If `warn_on_size` is true, logs a warning when the serialized policy exceeds 5KB.
/// Extracts `routing_preferences` from a JSON body, returning the cleaned body bytes
/// and the parsed preferences. The field is removed from the JSON before re-serializing
/// so downstream parsers don't see it.
pub fn extract_routing_policy(
raw_bytes: &[u8],
warn_on_size: bool,
) -> Result<(Bytes, Option<Vec<ModelUsagePreference>>), String> {
) -> Result<(Bytes, Option<Vec<TopLevelRoutingPreference>>), String> {
let mut json_body: serde_json::Value = serde_json::from_slice(raw_bytes)
.map_err(|err| format!("Failed to parse JSON: {}", err))?;
let preferences = json_body
let routing_preferences = json_body
.as_object_mut()
.and_then(|obj| obj.remove("routing_policy"))
.and_then(|policy_value| {
if warn_on_size {
let policy_str = serde_json::to_string(&policy_value).unwrap_or_default();
if policy_str.len() > ROUTING_POLICY_SIZE_WARNING_BYTES {
warn!(
size_bytes = policy_str.len(),
limit_bytes = ROUTING_POLICY_SIZE_WARNING_BYTES,
"routing_policy exceeds recommended size limit"
);
}
}
match serde_json::from_value::<Vec<ModelUsagePreference>>(policy_value) {
.and_then(|o| o.remove("routing_preferences"))
.and_then(
|value| match serde_json::from_value::<Vec<TopLevelRoutingPreference>>(value) {
Ok(prefs) => {
info!(
num_models = prefs.len(),
"using inline routing_policy from request body"
num_routes = prefs.len(),
"using inline routing_preferences from request body"
);
Some(prefs)
}
Err(err) => {
warn!(error = %err, "failed to parse routing_policy");
warn!(error = %err, "failed to parse routing_preferences");
None
}
}
});
},
);
let bytes = Bytes::from(serde_json::to_vec(&json_body).unwrap());
Ok((bytes, preferences))
Ok((bytes, routing_preferences))
}
#[derive(serde::Serialize)]
struct RoutingDecisionResponse {
model: String,
/// Ranked model list — use first, fall back to next on 429/5xx.
models: Vec<String>,
route: Option<String>,
trace_id: String,
}
@ -136,8 +122,9 @@ async fn routing_decision_inner(
"routing decision request body received"
);
// Extract routing_policy from request body before parsing as ProviderRequestType
let (chat_request_bytes, inline_preferences) = match extract_routing_policy(&raw_bytes, true) {
// Extract routing_preferences from body before parsing as ProviderRequestType
let (chat_request_bytes, inline_routing_preferences) = match extract_routing_policy(&raw_bytes)
{
Ok(result) => result,
Err(err) => {
warn!(error = %err, "failed to parse request JSON");
@ -164,27 +151,27 @@ async fn routing_decision_inner(
}
};
// Call the existing routing logic with inline preferences
let routing_result = router_chat_get_upstream_model(
router_service,
client_request,
&traceparent,
&request_path,
&request_id,
inline_preferences,
inline_routing_preferences,
)
.await;
match routing_result {
Ok(result) => {
let response = RoutingDecisionResponse {
model: result.model_name,
models: result.models,
route: result.route_name,
trace_id,
};
info!(
model = %response.model,
primary_model = %response.models.first().map(|s| s.as_str()).unwrap_or("none"),
total_models = response.models.len(),
route = ?response.route,
"routing decision completed"
);
@ -227,101 +214,70 @@ mod tests {
#[test]
fn extract_routing_policy_no_policy() {
let body = make_chat_body("");
let (cleaned, prefs) = extract_routing_policy(&body, false).unwrap();
let (cleaned, prefs) = extract_routing_policy(&body).unwrap();
assert!(prefs.is_none());
let cleaned_json: serde_json::Value = serde_json::from_slice(&cleaned).unwrap();
assert_eq!(cleaned_json["model"], "gpt-4o-mini");
assert!(cleaned_json.get("routing_policy").is_none());
}
#[test]
fn extract_routing_policy_valid_policy() {
let policy = r#""routing_policy": [
{
"model": "openai/gpt-4o",
"routing_preferences": [
{"name": "coding", "description": "code generation tasks"}
]
},
{
"model": "openai/gpt-4o-mini",
"routing_preferences": [
{"name": "general", "description": "general questions"}
]
}
]"#;
let body = make_chat_body(policy);
let (cleaned, prefs) = extract_routing_policy(&body, false).unwrap();
let prefs = prefs.expect("should have parsed preferences");
assert_eq!(prefs.len(), 2);
assert_eq!(prefs[0].model, "openai/gpt-4o");
assert_eq!(prefs[0].routing_preferences[0].name, "coding");
assert_eq!(prefs[1].model, "openai/gpt-4o-mini");
assert_eq!(prefs[1].routing_preferences[0].name, "general");
// routing_policy should be stripped from cleaned body
let cleaned_json: serde_json::Value = serde_json::from_slice(&cleaned).unwrap();
assert!(cleaned_json.get("routing_policy").is_none());
assert_eq!(cleaned_json["model"], "gpt-4o-mini");
}
#[test]
fn extract_routing_policy_invalid_policy_returns_none() {
// routing_policy is present but has wrong shape
let policy = r#""routing_policy": "not-an-array""#;
let body = make_chat_body(policy);
let (cleaned, prefs) = extract_routing_policy(&body, false).unwrap();
// Invalid policy should be ignored (returns None), not error
assert!(prefs.is_none());
// routing_policy should still be stripped from cleaned body
let cleaned_json: serde_json::Value = serde_json::from_slice(&cleaned).unwrap();
assert!(cleaned_json.get("routing_policy").is_none());
}
#[test]
fn extract_routing_policy_invalid_json_returns_error() {
let body = b"not valid json";
let result = extract_routing_policy(body, false);
let result = extract_routing_policy(body);
assert!(result.is_err());
assert!(result.unwrap_err().contains("Failed to parse JSON"));
}
#[test]
fn extract_routing_policy_empty_array() {
let policy = r#""routing_policy": []"#;
fn extract_routing_policy_routing_preferences() {
let policy = r#""routing_preferences": [
{
"name": "code generation",
"description": "generate new code",
"models": ["openai/gpt-4o", "openai/gpt-4o-mini"],
"selection_policy": {"prefer": "fastest"}
}
]"#;
let body = make_chat_body(policy);
let (_, prefs) = extract_routing_policy(&body, false).unwrap();
let (cleaned, prefs) = extract_routing_policy(&body).unwrap();
let prefs = prefs.expect("empty array is valid");
assert_eq!(prefs.len(), 0);
let prefs = prefs.expect("should have parsed routing_preferences");
assert_eq!(prefs.len(), 1);
assert_eq!(prefs[0].name, "code generation");
assert_eq!(prefs[0].models, vec!["openai/gpt-4o", "openai/gpt-4o-mini"]);
let cleaned_json: serde_json::Value = serde_json::from_slice(&cleaned).unwrap();
assert!(cleaned_json.get("routing_preferences").is_none());
}
#[test]
fn extract_routing_policy_preserves_other_fields() {
let policy = r#""routing_policy": [{"model": "gpt-4o", "routing_preferences": [{"name": "test", "description": "test"}]}], "temperature": 0.5, "max_tokens": 100"#;
let policy = r#""routing_preferences": [{"name": "test", "description": "test", "models": ["gpt-4o"], "selection_policy": {"prefer": "none"}}], "temperature": 0.5, "max_tokens": 100"#;
let body = make_chat_body(policy);
let (cleaned, prefs) = extract_routing_policy(&body, false).unwrap();
let (cleaned, prefs) = extract_routing_policy(&body).unwrap();
assert!(prefs.is_some());
let cleaned_json: serde_json::Value = serde_json::from_slice(&cleaned).unwrap();
assert_eq!(cleaned_json["temperature"], 0.5);
assert_eq!(cleaned_json["max_tokens"], 100);
assert!(cleaned_json.get("routing_policy").is_none());
assert!(cleaned_json.get("routing_preferences").is_none());
}
#[test]
fn routing_decision_response_serialization() {
let response = RoutingDecisionResponse {
model: "openai/gpt-4o".to_string(),
models: vec![
"openai/gpt-4o-mini".to_string(),
"openai/gpt-4o".to_string(),
],
route: Some("code_generation".to_string()),
trace_id: "abc123".to_string(),
};
let json = serde_json::to_string(&response).unwrap();
let parsed: serde_json::Value = serde_json::from_str(&json).unwrap();
assert_eq!(parsed["model"], "openai/gpt-4o");
assert_eq!(parsed["models"][0], "openai/gpt-4o-mini");
assert_eq!(parsed["models"][1], "openai/gpt-4o");
assert_eq!(parsed["route"], "code_generation");
assert_eq!(parsed["trace_id"], "abc123");
}
@ -329,13 +285,13 @@ mod tests {
#[test]
fn routing_decision_response_serialization_no_route() {
let response = RoutingDecisionResponse {
model: "none".to_string(),
models: vec!["none".to_string()],
route: None,
trace_id: "abc123".to_string(),
};
let json = serde_json::to_string(&response).unwrap();
let parsed: serde_json::Value = serde_json::from_str(&json).unwrap();
assert_eq!(parsed["model"], "none");
assert_eq!(parsed["models"][0], "none");
assert!(parsed["route"].is_null());
}
}

156
crates/brightstaff/src/main.rs
View file

@ -6,6 +6,7 @@ use brightstaff::handlers::llm::llm_chat;
use brightstaff::handlers::models::list_models;
use brightstaff::handlers::routing_service::routing_decision;
use brightstaff::router::llm::RouterService;
use brightstaff::router::model_metrics::ModelMetricsService;
use brightstaff::router::orchestrator::OrchestratorService;
use brightstaff::state::memory::MemoryConversationalStorage;
use brightstaff::state::postgresql::PostgreSQLConversationStorage;
@ -40,6 +41,17 @@ const DEFAULT_ROUTING_MODEL_NAME: &str = "Arch-Router";
const DEFAULT_ORCHESTRATOR_LLM_PROVIDER: &str = "plano-orchestrator";
const DEFAULT_ORCHESTRATOR_MODEL_NAME: &str = "Plano-Orchestrator";
/// Parse a version string like `v0.4.0`, `v0.3.0`, `0.2.0` into a `(major, minor, patch)` tuple.
/// Missing parts default to 0. Non-numeric parts are treated as 0.
fn parse_semver(version: &str) -> (u32, u32, u32) {
let v = version.trim_start_matches('v');
let mut parts = v.splitn(3, '.').map(|p| p.parse::<u32>().unwrap_or(0));
let major = parts.next().unwrap_or(0);
let minor = parts.next().unwrap_or(0);
let patch = parts.next().unwrap_or(0);
(major, minor, patch)
}
/// CORS pre-flight response for the models endpoint.
fn cors_preflight() -> Result<Response<BoxBody<Bytes, hyper::Error>>, hyper::Error> {
let mut response = Response::new(empty());
@ -162,8 +174,150 @@ async fn init_app_state(
.map(|p| p.name.clone())
.unwrap_or_else(|| DEFAULT_ROUTING_LLM_PROVIDER.to_string());
// Validate that top-level routing_preferences requires v0.4.0+.
let config_version = parse_semver(&config.version);
let is_v040_plus = config_version >= (0, 4, 0);
if !is_v040_plus && config.routing_preferences.is_some() {
return Err(
"top-level routing_preferences requires version v0.4.0 or above. \
Update the version field or remove routing_preferences."
.into(),
);
}
// Validate that all models referenced in top-level routing_preferences exist in model_providers.
// The CLI renders model_providers with `name` = "openai/gpt-4o" and `model` = "gpt-4o",
// so we accept a match against either field.
if let Some(ref route_prefs) = config.routing_preferences {
let provider_model_names: std::collections::HashSet<&str> = config
.model_providers
.iter()
.flat_map(|p| std::iter::once(p.name.as_str()).chain(p.model.as_deref()))
.collect();
for pref in route_prefs {
for model in &pref.models {
if !provider_model_names.contains(model.as_str()) {
return Err(format!(
"routing_preferences route '{}' references model '{}' \
which is not declared in model_providers",
pref.name, model
)
.into());
}
}
}
}
// Validate and initialize ModelMetricsService if model_metrics_sources is configured.
let metrics_service: Option<Arc<ModelMetricsService>> = if let Some(ref sources) =
config.model_metrics_sources
{
use common::configuration::MetricsSource;
let cost_count = sources
.iter()
.filter(|s| matches!(s, MetricsSource::CostMetrics { .. }))
.count();
let prom_count = sources
.iter()
.filter(|s| matches!(s, MetricsSource::PrometheusMetrics { .. }))
.count();
let do_count = sources
.iter()
.filter(|s| matches!(s, MetricsSource::DigitalOceanPricing { .. }))
.count();
if cost_count > 1 {
return Err("model_metrics_sources: only one cost_metrics source is allowed".into());
}
if prom_count > 1 {
return Err(
"model_metrics_sources: only one prometheus_metrics source is allowed".into(),
);
}
if do_count > 1 {
return Err(
"model_metrics_sources: only one digitalocean_pricing source is allowed".into(),
);
}
if cost_count > 0 && do_count > 0 {
return Err(
"model_metrics_sources: cost_metrics and digitalocean_pricing cannot both be configured — use one or the other".into(),
);
}
let svc = ModelMetricsService::new(sources, reqwest::Client::new()).await;
Some(Arc::new(svc))
} else {
None
};
// Validate that selection_policy.prefer is compatible with the configured metric sources.
if let Some(ref prefs) = config.routing_preferences {
use common::configuration::{MetricsSource, SelectionPreference};
let has_cost_source = config
.model_metrics_sources
.as_deref()
.unwrap_or_default()
.iter()
.any(|s| {
matches!(
s,
MetricsSource::CostMetrics { .. } | MetricsSource::DigitalOceanPricing { .. }
)
});
let has_prometheus = config
.model_metrics_sources
.as_deref()
.unwrap_or_default()
.iter()
.any(|s| matches!(s, MetricsSource::PrometheusMetrics { .. }));
for pref in prefs {
if pref.selection_policy.prefer == SelectionPreference::Cheapest && !has_cost_source {
return Err(format!(
"routing_preferences route '{}' uses prefer: cheapest but no cost data source is configured — \
add cost_metrics or digitalocean_pricing to model_metrics_sources",
pref.name
)
.into());
}
if pref.selection_policy.prefer == SelectionPreference::Fastest && !has_prometheus {
return Err(format!(
"routing_preferences route '{}' uses prefer: fastest but no prometheus_metrics source is configured — \
add prometheus_metrics to model_metrics_sources",
pref.name
)
.into());
}
}
}
// Warn about models in routing_preferences that have no matching pricing/latency data.
if let (Some(ref prefs), Some(ref svc)) = (&config.routing_preferences, &metrics_service) {
let cost_data = svc.cost_snapshot().await;
let latency_data = svc.latency_snapshot().await;
for pref in prefs {
use common::configuration::SelectionPreference;
for model in &pref.models {
let missing = match pref.selection_policy.prefer {
SelectionPreference::Cheapest => !cost_data.contains_key(model.as_str()),
SelectionPreference::Fastest => !latency_data.contains_key(model.as_str()),
_ => false,
};
if missing {
warn!(
model = %model,
route = %pref.name,
"model has no metric data — will be ranked last"
);
}
}
}
}
let router_service = Arc::new(RouterService::new(
config.model_providers.clone(),
config.routing_preferences.clone(),
metrics_service,
format!("{llm_provider_url}{CHAT_COMPLETIONS_PATH}"),
routing_model_name,
routing_llm_provider,

106
crates/brightstaff/src/router/llm.rs
View file

@ -1,15 +1,18 @@
use std::{collections::HashMap, sync::Arc};
use common::{
configuration::{LlmProvider, ModelUsagePreference, RoutingPreference},
configuration::TopLevelRoutingPreference,
consts::{ARCH_PROVIDER_HINT_HEADER, REQUEST_ID_HEADER, TRACE_PARENT_HEADER},
};
use super::router_model::{ModelUsagePreference, RoutingPreference};
use hermesllm::apis::openai::Message;
use hyper::header;
use thiserror::Error;
use tracing::{debug, info};
use super::http::{self, post_and_extract_content};
use super::model_metrics::ModelMetricsService;
use super::router_model::RouterModel;
use crate::router::router_model_v1;
@ -19,7 +22,8 @@ pub struct RouterService {
client: reqwest::Client,
router_model: Arc<dyn RouterModel>,
routing_provider_name: String,
llm_usage_defined: bool,
top_level_preferences: HashMap<String, TopLevelRoutingPreference>,
metrics_service: Option<Arc<ModelMetricsService>>,
}
#[derive(Debug, Error)]
@ -35,29 +39,37 @@ pub type Result<T> = std::result::Result<T, RoutingError>;
impl RouterService {
pub fn new(
providers: Vec<LlmProvider>,
top_level_prefs: Option<Vec<TopLevelRoutingPreference>>,
metrics_service: Option<Arc<ModelMetricsService>>,
router_url: String,
routing_model_name: String,
routing_provider_name: String,
) -> Self {
let providers_with_usage = providers
.iter()
.filter(|provider| provider.routing_preferences.is_some())
.cloned()
.collect::<Vec<LlmProvider>>();
let top_level_preferences: HashMap<String, TopLevelRoutingPreference> = top_level_prefs
.map_or_else(HashMap::new, |prefs| {
prefs.into_iter().map(|p| (p.name.clone(), p)).collect()
});
let llm_routes: HashMap<String, Vec<RoutingPreference>> = providers_with_usage
// Build sentinel routes for RouterModelV1: route_name → first model.
// RouterModelV1 uses this to build its prompt; RouterService overrides
// the model selection via rank_models() after the route is determined.
let sentinel_routes: HashMap<String, Vec<RoutingPreference>> = top_level_preferences
.iter()
.filter_map(|provider| {
provider
.routing_preferences
.as_ref()
.map(|prefs| (provider.name.clone(), prefs.clone()))
.filter_map(|(name, pref)| {
pref.models.first().map(|first_model| {
(
first_model.clone(),
vec![RoutingPreference {
name: name.clone(),
description: pref.description.clone(),
}],
)
})
})
.collect();
let router_model = Arc::new(router_model_v1::RouterModelV1::new(
llm_routes,
sentinel_routes,
routing_model_name,
router_model_v1::MAX_TOKEN_LEN,
));
@ -67,7 +79,8 @@ impl RouterService {
client: reqwest::Client::new(),
router_model,
routing_provider_name,
llm_usage_defined: !providers_with_usage.is_empty(),
top_level_preferences,
metrics_service,
}
}
@ -75,24 +88,43 @@ impl RouterService {
&self,
messages: &[Message],
traceparent: &str,
usage_preferences: Option<Vec<ModelUsagePreference>>,
inline_routing_preferences: Option<Vec<TopLevelRoutingPreference>>,
request_id: &str,
) -> Result<Option<(String, String)>> {
) -> Result<Option<(String, Vec<String>)>> {
if messages.is_empty() {
return Ok(None);
}
if usage_preferences
.as_ref()
.is_none_or(|prefs| prefs.len() < 2)
&& !self.llm_usage_defined
{
// Build inline top-level map from request if present (inline overrides config).
let inline_top_map: Option<HashMap<String, TopLevelRoutingPreference>> =
inline_routing_preferences
.map(|prefs| prefs.into_iter().map(|p| (p.name.clone(), p)).collect());
// No routing defined — skip the router call entirely.
if inline_top_map.is_none() && self.top_level_preferences.is_empty() {
return Ok(None);
}
// For inline overrides, build synthetic ModelUsagePreference list so RouterModelV1
// generates the correct prompt (route name + description pairs).
// For config-level prefs the sentinel routes are already baked into RouterModelV1.
let effective_usage_preferences: Option<Vec<ModelUsagePreference>> =
inline_top_map.as_ref().map(|inline_map| {
inline_map
.values()
.map(|p| ModelUsagePreference {
model: p.models.first().cloned().unwrap_or_default(),
routing_preferences: vec![RoutingPreference {
name: p.name.clone(),
description: p.description.clone(),
}],
})
.collect()
});
let router_request = self
.router_model
.generate_request(messages, &usage_preferences);
.generate_request(messages, &effective_usage_preferences);
debug!(
model = %self.router_model.get_model_name(),
@ -132,17 +164,37 @@ impl RouterService {
return Ok(None);
};
// Parse the route name from the router response.
let parsed = self
.router_model
.parse_response(&content, &usage_preferences)?;
.parse_response(&content, &effective_usage_preferences)?;
let result = if let Some((route_name, _sentinel)) = parsed {
let top_pref = inline_top_map
.as_ref()
.and_then(|m| m.get(&route_name))
.or_else(|| self.top_level_preferences.get(&route_name));
if let Some(pref) = top_pref {
let ranked = match &self.metrics_service {
Some(svc) => svc.rank_models(&pref.models, &pref.selection_policy).await,
None => pref.models.clone(),
};
Some((route_name, ranked))
} else {
None
}
} else {
None
};
info!(
content = %content.replace("\n", "\\n"),
selected_model = ?parsed,
selected_model = ?result,
response_time_ms = elapsed.as_millis(),
"arch-router determined route"
);
Ok(parsed)
Ok(result)
}
}

1
crates/brightstaff/src/router/mod.rs
View file

@ -1,5 +1,6 @@
pub(crate) mod http;
pub mod llm;
pub mod model_metrics;
pub mod orchestrator;
pub mod orchestrator_model;
pub mod orchestrator_model_v1;

419
crates/brightstaff/src/router/model_metrics.rs Normal file
View file

@ -0,0 +1,419 @@
use std::collections::HashMap;
use std::sync::Arc;
use std::time::Duration;
use common::configuration::{MetricsSource, SelectionPolicy, SelectionPreference};
use tokio::sync::RwLock;
use tracing::{info, warn};
const DO_PRICING_URL: &str = "https://api.digitalocean.com/v2/gen-ai/models/catalog";
pub struct ModelMetricsService {
cost: Arc<RwLock<HashMap<String, f64>>>,
latency: Arc<RwLock<HashMap<String, f64>>>,
}
impl ModelMetricsService {
pub async fn new(sources: &[MetricsSource], client: reqwest::Client) -> Self {
let cost_data = Arc::new(RwLock::new(HashMap::new()));
let latency_data = Arc::new(RwLock::new(HashMap::new()));
for source in sources {
match source {
MetricsSource::CostMetrics {
url,
refresh_interval,
auth,
} => {
let data = fetch_cost_metrics(url, auth.as_ref(), &client).await;
info!(models = data.len(), url = %url, "fetched cost metrics");
*cost_data.write().await = data;
if let Some(interval_secs) = refresh_interval {
let cost_clone = Arc::clone(&cost_data);
let client_clone = client.clone();
let url = url.clone();
let auth = auth.clone();
let interval = Duration::from_secs(*interval_secs);
tokio::spawn(async move {
loop {
tokio::time::sleep(interval).await;
let data =
fetch_cost_metrics(&url, auth.as_ref(), &client_clone).await;
info!(models = data.len(), url = %url, "refreshed cost metrics");
*cost_clone.write().await = data;
}
});
}
}
MetricsSource::PrometheusMetrics {
url,
query,
refresh_interval,
} => {
let data = fetch_prometheus_metrics(url, query, &client).await;
info!(models = data.len(), url = %url, "fetched prometheus latency metrics");
*latency_data.write().await = data;
if let Some(interval_secs) = refresh_interval {
let latency_clone = Arc::clone(&latency_data);
let client_clone = client.clone();
let url = url.clone();
let query = query.clone();
let interval = Duration::from_secs(*interval_secs);
tokio::spawn(async move {
loop {
tokio::time::sleep(interval).await;
let data =
fetch_prometheus_metrics(&url, &query, &client_clone).await;
info!(models = data.len(), url = %url, "refreshed prometheus latency metrics");
*latency_clone.write().await = data;
}
});
}
}
MetricsSource::DigitalOceanPricing {
refresh_interval,
model_aliases,
} => {
let aliases = model_aliases.clone().unwrap_or_default();
let data = fetch_do_pricing(&client, &aliases).await;
info!(models = data.len(), "fetched digitalocean pricing");
*cost_data.write().await = data;
if let Some(interval_secs) = refresh_interval {
let cost_clone = Arc::clone(&cost_data);
let client_clone = client.clone();
let interval = Duration::from_secs(*interval_secs);
tokio::spawn(async move {
loop {
tokio::time::sleep(interval).await;
let data = fetch_do_pricing(&client_clone, &aliases).await;
info!(models = data.len(), "refreshed digitalocean pricing");
*cost_clone.write().await = data;
}
});
}
}
}
}
ModelMetricsService {
cost: cost_data,
latency: latency_data,
}
}
/// Rank `models` by `policy`, returning them in preference order.
/// Models with no metric data are appended at the end in their original order.
pub async fn rank_models(&self, models: &[String], policy: &SelectionPolicy) -> Vec<String> {
match policy.prefer {
SelectionPreference::Cheapest => {
let data = self.cost.read().await;
for m in models {
if !data.contains_key(m.as_str()) {
warn!(model = %m, "no cost data for model — ranking last (prefer: cheapest)");
}
}
rank_by_ascending_metric(models, &data)
}
SelectionPreference::Fastest => {
let data = self.latency.read().await;
for m in models {
if !data.contains_key(m.as_str()) {
warn!(model = %m, "no latency data for model — ranking last (prefer: fastest)");
}
}
rank_by_ascending_metric(models, &data)
}
SelectionPreference::None => models.to_vec(),
}
}
/// Returns a snapshot of the current cost data. Used at startup to warn about unmatched models.
pub async fn cost_snapshot(&self) -> HashMap<String, f64> {
self.cost.read().await.clone()
}
/// Returns a snapshot of the current latency data. Used at startup to warn about unmatched models.
pub async fn latency_snapshot(&self) -> HashMap<String, f64> {
self.latency.read().await.clone()
}
}
fn rank_by_ascending_metric(models: &[String], data: &HashMap<String, f64>) -> Vec<String> {
let mut with_data: Vec<(&String, f64)> = models
.iter()
.filter_map(|m| data.get(m.as_str()).map(|v| (m, *v)))
.collect();
with_data.sort_by(|a, b| a.1.partial_cmp(&b.1).unwrap_or(std::cmp::Ordering::Equal));
let without_data: Vec<&String> = models
.iter()
.filter(|m| !data.contains_key(m.as_str()))
.collect();
with_data
.iter()
.map(|(m, _)| (*m).clone())
.chain(without_data.iter().map(|m| (*m).clone()))
.collect()
}
#[derive(serde::Deserialize)]
struct CostEntry {
input_per_million: f64,
output_per_million: f64,
}
async fn fetch_cost_metrics(
url: &str,
auth: Option<&common::configuration::MetricsAuth>,
client: &reqwest::Client,
) -> HashMap<String, f64> {
let mut req = client.get(url);
if let Some(auth) = auth {
if auth.auth_type == "bearer" {
req = req.header("Authorization", format!("Bearer {}", auth.token));
} else {
warn!(auth_type = %auth.auth_type, "unsupported auth type for cost_metrics, skipping auth");
}
}
match req.send().await {
Ok(resp) => match resp.json::<HashMap<String, CostEntry>>().await {
Ok(data) => data
.into_iter()
.map(|(k, v)| (k, v.input_per_million + v.output_per_million))
.collect(),
Err(err) => {
warn!(error = %err, url = %url, "failed to parse cost metrics response");
HashMap::new()
}
},
Err(err) => {
warn!(error = %err, url = %url, "failed to fetch cost metrics");
HashMap::new()
}
}
}
#[derive(serde::Deserialize)]
struct DoModelList {
data: Vec<DoModel>,
}
#[derive(serde::Deserialize)]
struct DoModel {
model_id: String,
pricing: Option<DoPricing>,
}
#[derive(serde::Deserialize)]
struct DoPricing {
input_price_per_million: Option<f64>,
output_price_per_million: Option<f64>,
}
async fn fetch_do_pricing(
client: &reqwest::Client,
aliases: &HashMap<String, String>,
) -> HashMap<String, f64> {
match client.get(DO_PRICING_URL).send().await {
Ok(resp) => match resp.json::<DoModelList>().await {
Ok(list) => list
.data
.into_iter()
.filter_map(|m| {
let pricing = m.pricing?;
let raw_key = m.model_id.clone();
let key = aliases.get(&raw_key).cloned().unwrap_or(raw_key);
let cost = pricing.input_price_per_million.unwrap_or(0.0)
+ pricing.output_price_per_million.unwrap_or(0.0);
Some((key, cost))
})
.collect(),
Err(err) => {
warn!(error = %err, url = DO_PRICING_URL, "failed to parse digitalocean pricing response");
HashMap::new()
}
},
Err(err) => {
warn!(error = %err, url = DO_PRICING_URL, "failed to fetch digitalocean pricing");
HashMap::new()
}
}
}
#[derive(serde::Deserialize)]
struct PrometheusResponse {
data: PrometheusData,
}
#[derive(serde::Deserialize)]
struct PrometheusData {
result: Vec<PrometheusResult>,
}
#[derive(serde::Deserialize)]
struct PrometheusResult {
metric: HashMap<String, String>,
value: (f64, String), // (timestamp, value_str)
}
async fn fetch_prometheus_metrics(
url: &str,
query: &str,
client: &reqwest::Client,
) -> HashMap<String, f64> {
let query_url = format!("{}/api/v1/query", url.trim_end_matches('/'));
match client
.get(&query_url)
.query(&[("query", query)])
.send()
.await
{
Ok(resp) => match resp.json::<PrometheusResponse>().await {
Ok(prom) => prom
.data
.result
.into_iter()
.filter_map(|r| {
let model_name = r.metric.get("model_name")?.clone();
let value: f64 = r.value.1.parse().ok()?;
Some((model_name, value))
})
.collect(),
Err(err) => {
warn!(error = %err, url = %query_url, "failed to parse prometheus response");
HashMap::new()
}
},
Err(err) => {
warn!(error = %err, url = %query_url, "failed to fetch prometheus metrics");
HashMap::new()
}
}
}
#[cfg(test)]
mod tests {
use super::*;
use common::configuration::SelectionPreference;
fn make_policy(prefer: SelectionPreference) -> SelectionPolicy {
SelectionPolicy { prefer }
}
#[test]
fn test_rank_by_ascending_metric_picks_lowest_first() {
let models = vec!["a".to_string(), "b".to_string(), "c".to_string()];
let mut data = HashMap::new();
data.insert("a".to_string(), 0.01);
data.insert("b".to_string(), 0.005);
data.insert("c".to_string(), 0.02);
assert_eq!(
rank_by_ascending_metric(&models, &data),
vec!["b", "a", "c"]
);
}
#[test]
fn test_rank_by_ascending_metric_no_data_preserves_order() {
let models = vec!["x".to_string(), "y".to_string()];
let data = HashMap::new();
assert_eq!(rank_by_ascending_metric(&models, &data), vec!["x", "y"]);
}
#[test]
fn test_rank_by_ascending_metric_partial_data() {
let models = vec!["a".to_string(), "b".to_string()];
let mut data = HashMap::new();
data.insert("b".to_string(), 100.0);
assert_eq!(rank_by_ascending_metric(&models, &data), vec!["b", "a"]);
}
#[tokio::test]
async fn test_rank_models_cheapest() {
let service = ModelMetricsService {
cost: Arc::new(RwLock::new({
let mut m = HashMap::new();
m.insert("gpt-4o".to_string(), 0.005);
m.insert("gpt-4o-mini".to_string(), 0.0001);
m
})),
latency: Arc::new(RwLock::new(HashMap::new())),
};
let models = vec!["gpt-4o".to_string(), "gpt-4o-mini".to_string()];
let result = service
.rank_models(&models, &make_policy(SelectionPreference::Cheapest))
.await;
assert_eq!(result, vec!["gpt-4o-mini", "gpt-4o"]);
}
#[tokio::test]
async fn test_rank_models_fastest() {
let service = ModelMetricsService {
cost: Arc::new(RwLock::new(HashMap::new())),
latency: Arc::new(RwLock::new({
let mut m = HashMap::new();
m.insert("gpt-4o".to_string(), 200.0);
m.insert("claude-sonnet".to_string(), 120.0);
m
})),
};
let models = vec!["gpt-4o".to_string(), "claude-sonnet".to_string()];
let result = service
.rank_models(&models, &make_policy(SelectionPreference::Fastest))
.await;
assert_eq!(result, vec!["claude-sonnet", "gpt-4o"]);
}
#[tokio::test]
async fn test_rank_models_fallback_no_metrics() {
let service = ModelMetricsService {
cost: Arc::new(RwLock::new(HashMap::new())),
latency: Arc::new(RwLock::new(HashMap::new())),
};
let models = vec!["model-a".to_string(), "model-b".to_string()];
let result = service
.rank_models(&models, &make_policy(SelectionPreference::Cheapest))
.await;
assert_eq!(result, vec!["model-a", "model-b"]);
}
#[tokio::test]
async fn test_rank_models_partial_data_appended_last() {
let service = ModelMetricsService {
cost: Arc::new(RwLock::new({
let mut m = HashMap::new();
m.insert("gpt-4o".to_string(), 0.005);
m
})),
latency: Arc::new(RwLock::new(HashMap::new())),
};
let models = vec!["gpt-4o-mini".to_string(), "gpt-4o".to_string()];
let result = service
.rank_models(&models, &make_policy(SelectionPreference::Cheapest))
.await;
assert_eq!(result, vec!["gpt-4o", "gpt-4o-mini"]);
}
#[tokio::test]
async fn test_rank_models_none_preserves_order() {
let service = ModelMetricsService {
cost: Arc::new(RwLock::new({
let mut m = HashMap::new();
m.insert("gpt-4o-mini".to_string(), 0.0001);
m.insert("gpt-4o".to_string(), 0.005);
m
})),
latency: Arc::new(RwLock::new(HashMap::new())),
};
let models = vec!["gpt-4o".to_string(), "gpt-4o-mini".to_string()];
let result = service
.rank_models(&models, &make_policy(SelectionPreference::None))
.await;
// none → original order, despite gpt-4o-mini being cheaper
assert_eq!(result, vec!["gpt-4o", "gpt-4o-mini"]);
}
}

16
crates/brightstaff/src/router/router_model.rs
View file

@ -1,5 +1,5 @@
use common::configuration::ModelUsagePreference;
use hermesllm::apis::openai::{ChatCompletionsRequest, Message};
use serde::{Deserialize, Serialize};
use thiserror::Error;
#[derive(Debug, Error)]
@ -10,6 +10,20 @@ pub enum RoutingModelError {
pub type Result<T> = std::result::Result<T, RoutingModelError>;
/// Internal route descriptor passed to the router model to build its prompt.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct RoutingPreference {
pub name: String,
pub description: String,
}
/// Groups a model with its routing preferences (used internally by RouterModelV1).
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ModelUsagePreference {
pub model: String,
pub routing_preferences: Vec<RoutingPreference>,
}
pub trait RouterModel: Send + Sync {
fn generate_request(
&self,

2
crates/brightstaff/src/router/router_model_v1.rs
View file

@ -1,6 +1,6 @@
use std::collections::HashMap;
use common::configuration::{ModelUsagePreference, RoutingPreference};
use super::router_model::{ModelUsagePreference, RoutingPreference};
use hermesllm::apis::openai::{ChatCompletionsRequest, Message, MessageContent, Role};
use hermesllm::transforms::lib::ExtractText;
use serde::{Deserialize, Serialize};