mirror of
https://github.com/katanemo/plano.git
synced 2026-06-02 14:35:14 +02:00
plano orchestration using plano orchestration 4b model (#637)
This commit is contained in:
Adil Hafeez
GitHub
parent
60162e0575
commit
15fbb6c3af
40 changed files with 4054 additions and 449 deletions
|
|
@ -17,7 +17,7 @@ use tracing::{debug, info, warn};
|
|||
use super::agent_selector::{AgentSelectionError, AgentSelector};
|
||||
use super::pipeline_processor::{PipelineError, PipelineProcessor};
|
||||
use super::response_handler::ResponseHandler;
|
||||
use crate::router::llm_router::RouterService;
|
||||
use crate::router::plano_orchestrator::OrchestratorService;
|
||||
use crate::tracing::{OperationNameBuilder, operation_component, http};
|
||||
|
||||
/// Main errors for agent chat completions
|
||||
|
|
@ -37,7 +37,7 @@ pub enum AgentFilterChainError {
|
|||
|
||||
pub async fn agent_chat(
|
||||
request: Request<hyper::body::Incoming>,
|
||||
router_service: Arc<RouterService>,
|
||||
orchestrator_service: Arc<OrchestratorService>,
|
||||
_: String,
|
||||
agents_list: Arc<tokio::sync::RwLock<Option<Vec<common::configuration::Agent>>>>,
|
||||
listeners: Arc<tokio::sync::RwLock<Vec<common::configuration::Listener>>>,
|
||||
|
|
@ -45,7 +45,7 @@ pub async fn agent_chat(
|
|||
) -> Result<Response<BoxBody<Bytes, hyper::Error>>, hyper::Error> {
|
||||
match handle_agent_chat(
|
||||
request,
|
||||
router_service,
|
||||
orchestrator_service,
|
||||
agents_list,
|
||||
listeners,
|
||||
trace_collector,
|
||||
|
|
@ -123,13 +123,13 @@ pub async fn agent_chat(
|
|||
|
||||
async fn handle_agent_chat(
|
||||
request: Request<hyper::body::Incoming>,
|
||||
router_service: Arc<RouterService>,
|
||||
orchestrator_service: Arc<OrchestratorService>,
|
||||
agents_list: Arc<tokio::sync::RwLock<Option<Vec<common::configuration::Agent>>>>,
|
||||
listeners: Arc<tokio::sync::RwLock<Vec<common::configuration::Listener>>>,
|
||||
trace_collector: Arc<common::traces::TraceCollector>,
|
||||
) -> Result<Response<BoxBody<Bytes, hyper::Error>>, AgentFilterChainError> {
|
||||
// Initialize services
|
||||
let agent_selector = AgentSelector::new(router_service);
|
||||
let agent_selector = AgentSelector::new(orchestrator_service);
|
||||
let mut pipeline_processor = PipelineProcessor::default();
|
||||
let response_handler = ResponseHandler::new();
|
||||
|
||||
|
|
@ -186,15 +186,6 @@ async fn handle_agent_chat(
|
|||
|
||||
let message: Vec<OpenAIMessage> = client_request.get_messages();
|
||||
|
||||
// let chat_completions_request: ChatCompletionsRequest =
|
||||
// serde_json::from_slice(&chat_request_bytes).map_err(|err| {
|
||||
// warn!(
|
||||
// "Failed to parse request body as ChatCompletionsRequest: {}",
|
||||
// err
|
||||
// );
|
||||
// AgentFilterChainError::RequestParsing(err)
|
||||
// })?;
|
||||
|
||||
// Extract trace parent for routing
|
||||
let trace_parent = request_headers
|
||||
.iter()
|
||||
|
|
@ -215,94 +206,166 @@ async fn handle_agent_chat(
|
|||
(String::new(), None)
|
||||
};
|
||||
|
||||
// Select appropriate agent using arch router llm model
|
||||
let selected_agent = agent_selector
|
||||
.select_agent(&message, &listener, trace_parent.clone())
|
||||
// Select appropriate agents using arch orchestrator llm model
|
||||
let selection_span_id = generate_random_span_id();
|
||||
let selection_start_time = SystemTime::now();
|
||||
let selection_start_instant = Instant::now();
|
||||
|
||||
let selected_agents = agent_selector
|
||||
.select_agents(&message, &listener, trace_parent.clone())
|
||||
.await?;
|
||||
|
||||
debug!("Processing agent pipeline: {}", selected_agent.id);
|
||||
|
||||
// Record the start time for agent span
|
||||
let agent_start_time = SystemTime::now();
|
||||
let agent_start_instant = Instant::now();
|
||||
// let (span_id, trace_id) = trace_collector.start_span(
|
||||
// trace_parent.clone(),
|
||||
// operation_component::AGENT,
|
||||
// &format!("/agents{}", request_path),
|
||||
// &selected_agent.id,
|
||||
// );
|
||||
|
||||
let span_id = generate_random_span_id();
|
||||
|
||||
// Process the filter chain
|
||||
let chat_history = pipeline_processor
|
||||
.process_filter_chain(
|
||||
&message,
|
||||
&selected_agent,
|
||||
&agent_map,
|
||||
&request_headers,
|
||||
Some(&trace_collector),
|
||||
trace_id.clone(),
|
||||
span_id.clone(),
|
||||
)
|
||||
.await?;
|
||||
|
||||
// Get terminal agent and send final response
|
||||
let terminal_agent_name = selected_agent.id.clone();
|
||||
let terminal_agent = agent_map.get(&terminal_agent_name).unwrap();
|
||||
|
||||
debug!("Processing terminal agent: {}", terminal_agent_name);
|
||||
debug!("Terminal agent details: {:?}", terminal_agent);
|
||||
|
||||
let llm_response = pipeline_processor
|
||||
.invoke_agent(
|
||||
&chat_history,
|
||||
client_request,
|
||||
terminal_agent,
|
||||
&request_headers,
|
||||
trace_id.clone(),
|
||||
span_id.clone(),
|
||||
)
|
||||
.await?;
|
||||
|
||||
// Record agent span after processing is complete
|
||||
let agent_end_time = SystemTime::now();
|
||||
let agent_elapsed = agent_start_instant.elapsed();
|
||||
|
||||
// Build full path with /agents prefix
|
||||
let full_path = format!("/agents{}", request_path);
|
||||
|
||||
// Build operation name: POST {full_path} {agent_name}
|
||||
let operation_name = OperationNameBuilder::new()
|
||||
// Record agent selection span
|
||||
let selection_end_time = SystemTime::now();
|
||||
let selection_elapsed = selection_start_instant.elapsed();
|
||||
let selection_operation_name = OperationNameBuilder::new()
|
||||
.with_method("POST")
|
||||
.with_path(&full_path)
|
||||
.with_target(&terminal_agent_name)
|
||||
.with_path("/agents/select")
|
||||
.with_target(&listener.name)
|
||||
.build();
|
||||
|
||||
let mut span_builder = SpanBuilder::new(&operation_name)
|
||||
.with_span_id(span_id)
|
||||
let mut selection_span_builder = SpanBuilder::new(&selection_operation_name)
|
||||
.with_span_id(selection_span_id)
|
||||
.with_kind(SpanKind::Internal)
|
||||
.with_start_time(agent_start_time)
|
||||
.with_end_time(agent_end_time)
|
||||
.with_start_time(selection_start_time)
|
||||
.with_end_time(selection_end_time)
|
||||
.with_attribute(http::METHOD, "POST")
|
||||
.with_attribute(http::TARGET, full_path)
|
||||
.with_attribute("agent.name", terminal_agent_name.clone())
|
||||
.with_attribute("duration_ms", format!("{:.2}", agent_elapsed.as_secs_f64() * 1000.0));
|
||||
.with_attribute(http::TARGET, "/agents/select")
|
||||
.with_attribute("selection.listener", listener.name.clone())
|
||||
.with_attribute("selection.agent_count", selected_agents.len().to_string())
|
||||
.with_attribute("selection.agents", selected_agents.iter().map(|a| a.id.as_str()).collect::<Vec<_>>().join(","))
|
||||
.with_attribute("duration_ms", format!("{:.2}", selection_elapsed.as_secs_f64() * 1000.0));
|
||||
|
||||
if !trace_id.is_empty() {
|
||||
span_builder = span_builder.with_trace_id(trace_id);
|
||||
selection_span_builder = selection_span_builder.with_trace_id(trace_id.clone());
|
||||
}
|
||||
if let Some(parent_id) = parent_span_id {
|
||||
span_builder = span_builder.with_parent_span_id(parent_id);
|
||||
if let Some(parent_id) = parent_span_id.clone() {
|
||||
selection_span_builder = selection_span_builder.with_parent_span_id(parent_id);
|
||||
}
|
||||
|
||||
let span = span_builder.build();
|
||||
// Use plano(agent) as service name for the agent processing span
|
||||
trace_collector.record_span(operation_component::AGENT, span);
|
||||
let selection_span = selection_span_builder.build();
|
||||
trace_collector.record_span(operation_component::ORCHESTRATOR, selection_span);
|
||||
|
||||
// Create streaming response
|
||||
response_handler
|
||||
.create_streaming_response(llm_response)
|
||||
.await
|
||||
.map_err(AgentFilterChainError::from)
|
||||
info!("Selected {} agent(s) for execution", selected_agents.len());
|
||||
|
||||
// Execute agents sequentially, passing output from one to the next
|
||||
let mut current_messages = message.clone();
|
||||
let agent_count = selected_agents.len();
|
||||
|
||||
for (agent_index, selected_agent) in selected_agents.iter().enumerate() {
|
||||
let is_last_agent = agent_index == agent_count - 1;
|
||||
|
||||
debug!(
|
||||
"Processing agent {}/{}: {}",
|
||||
agent_index + 1,
|
||||
agent_count,
|
||||
selected_agent.id
|
||||
);
|
||||
|
||||
// Record the start time for agent span
|
||||
let agent_start_time = SystemTime::now();
|
||||
let agent_start_instant = Instant::now();
|
||||
let span_id = generate_random_span_id();
|
||||
|
||||
// Get agent name
|
||||
let agent_name = selected_agent.id.clone();
|
||||
|
||||
// Process the filter chain
|
||||
let chat_history = pipeline_processor
|
||||
.process_filter_chain(
|
||||
¤t_messages,
|
||||
selected_agent,
|
||||
&agent_map,
|
||||
&request_headers,
|
||||
Some(&trace_collector),
|
||||
trace_id.clone(),
|
||||
span_id.clone(),
|
||||
)
|
||||
.await?;
|
||||
|
||||
// Get agent details and invoke
|
||||
let agent = agent_map.get(&agent_name).unwrap();
|
||||
|
||||
debug!("Invoking agent: {}", agent_name);
|
||||
|
||||
let llm_response = pipeline_processor
|
||||
.invoke_agent(
|
||||
&chat_history,
|
||||
client_request.clone(),
|
||||
agent,
|
||||
&request_headers,
|
||||
trace_id.clone(),
|
||||
span_id.clone(),
|
||||
)
|
||||
.await?;
|
||||
|
||||
// Record agent span
|
||||
let agent_end_time = SystemTime::now();
|
||||
let agent_elapsed = agent_start_instant.elapsed();
|
||||
let full_path = format!("/agents{}", request_path);
|
||||
let operation_name = OperationNameBuilder::new()
|
||||
.with_method("POST")
|
||||
.with_path(&full_path)
|
||||
.with_target(&agent_name)
|
||||
.build();
|
||||
|
||||
let mut span_builder = SpanBuilder::new(&operation_name)
|
||||
.with_span_id(span_id)
|
||||
.with_kind(SpanKind::Internal)
|
||||
.with_start_time(agent_start_time)
|
||||
.with_end_time(agent_end_time)
|
||||
.with_attribute(http::METHOD, "POST")
|
||||
.with_attribute(http::TARGET, full_path)
|
||||
.with_attribute("agent.name", agent_name.clone())
|
||||
.with_attribute("agent.sequence", format!("{}/{}", agent_index + 1, agent_count))
|
||||
.with_attribute("duration_ms", format!("{:.2}", agent_elapsed.as_secs_f64() * 1000.0));
|
||||
|
||||
if !trace_id.is_empty() {
|
||||
span_builder = span_builder.with_trace_id(trace_id.clone());
|
||||
}
|
||||
if let Some(parent_id) = parent_span_id.clone() {
|
||||
span_builder = span_builder.with_parent_span_id(parent_id);
|
||||
}
|
||||
|
||||
let span = span_builder.build();
|
||||
trace_collector.record_span(operation_component::AGENT, span);
|
||||
|
||||
// If this is the last agent, return the streaming response
|
||||
if is_last_agent {
|
||||
info!("Completed agent chain, returning response from last agent: {}", agent_name);
|
||||
return response_handler
|
||||
.create_streaming_response(llm_response)
|
||||
.await
|
||||
.map_err(AgentFilterChainError::from);
|
||||
}
|
||||
|
||||
// For intermediate agents, collect the full response and pass to next agent
|
||||
debug!("Collecting response from intermediate agent: {}", agent_name);
|
||||
let response_text = response_handler.collect_full_response(llm_response).await?;
|
||||
|
||||
info!(
|
||||
"Agent {} completed, passing {} character response to next agent",
|
||||
agent_name,
|
||||
response_text.len()
|
||||
);
|
||||
|
||||
// remove last message and add new one at the end
|
||||
let last_message = current_messages.pop().unwrap();
|
||||
|
||||
// Create a new message with the agent's response as assistant message
|
||||
// and add it to the conversation history
|
||||
current_messages.push(OpenAIMessage {
|
||||
role: hermesllm::apis::openai::Role::Assistant,
|
||||
content: hermesllm::apis::openai::MessageContent::Text(response_text),
|
||||
name: Some(agent_name.clone()),
|
||||
tool_calls: None,
|
||||
tool_call_id: None,
|
||||
});
|
||||
|
||||
current_messages.push(last_message);
|
||||
|
||||
}
|
||||
|
||||
// This should never be reached since we return in the last agent iteration
|
||||
unreachable!("Agent execution loop should have returned a response")
|
||||
}
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue