mirror of
https://github.com/katanemo/plano.git
synced 2026-05-07 23:02:43 +02:00
Add support for json based content types in Message (#480)
This commit is contained in:
parent
f5e77bbe65
commit
218e9c540d
16 changed files with 314 additions and 121 deletions
|
|
@ -1,14 +1,13 @@
|
|||
use std::sync::Arc;
|
||||
|
||||
use common::{
|
||||
api::open_ai::{ChatCompletionsResponse, Message},
|
||||
api::open_ai::{ChatCompletionsResponse, ContentType, Message},
|
||||
configuration::LlmProvider,
|
||||
consts::ARCH_PROVIDER_HINT_HEADER,
|
||||
utils::shorten_string,
|
||||
};
|
||||
use hyper::header;
|
||||
use thiserror::Error;
|
||||
use tracing::{info, warn};
|
||||
use tracing::{debug, info, warn};
|
||||
|
||||
use super::router_model::RouterModel;
|
||||
|
||||
|
|
@ -59,9 +58,9 @@ impl RouterService {
|
|||
.collect::<Vec<String>>()
|
||||
.join("\n");
|
||||
|
||||
info!(
|
||||
debug!(
|
||||
"llm_providers from config with usage: {}...",
|
||||
shorten_string(&llm_providers_with_usage_yaml.replace("\n", "\\n"))
|
||||
llm_providers_with_usage_yaml.replace("\n", "\\n")
|
||||
);
|
||||
|
||||
let router_model = Arc::new(super::router_model_v1::RouterModelV1::new(
|
||||
|
|
@ -83,7 +82,6 @@ impl RouterService {
|
|||
messages: &[Message],
|
||||
trace_parent: Option<String>,
|
||||
) -> Result<Option<String>> {
|
||||
|
||||
if !self.llm_usage_defined {
|
||||
return Ok(None);
|
||||
}
|
||||
|
|
@ -91,8 +89,14 @@ impl RouterService {
|
|||
let router_request = self.router_model.generate_request(messages);
|
||||
|
||||
info!(
|
||||
"router_request: {}",
|
||||
shorten_string(&serde_json::to_string(&router_request).unwrap()),
|
||||
"sending request to arch-router model: {}, endpoint: {}",
|
||||
self.router_model.get_model_name(),
|
||||
self.router_url
|
||||
);
|
||||
|
||||
debug!(
|
||||
"arch request body: {}",
|
||||
&serde_json::to_string(&router_request).unwrap(),
|
||||
);
|
||||
|
||||
let mut llm_route_request_headers = header::HeaderMap::new();
|
||||
|
|
@ -113,6 +117,7 @@ impl RouterService {
|
|||
);
|
||||
}
|
||||
|
||||
let start_time = std::time::Instant::now();
|
||||
let res = self
|
||||
.client
|
||||
.post(&self.router_url)
|
||||
|
|
@ -122,6 +127,7 @@ impl RouterService {
|
|||
.await?;
|
||||
|
||||
let body = res.text().await?;
|
||||
let router_response_time = start_time.elapsed();
|
||||
|
||||
let chat_completion_response: ChatCompletionsResponse = match serde_json::from_str(&body) {
|
||||
Ok(response) => response,
|
||||
|
|
@ -138,14 +144,18 @@ impl RouterService {
|
|||
}
|
||||
};
|
||||
|
||||
let selected_llm = self.router_model.parse_response(
|
||||
chat_completion_response.choices[0]
|
||||
.message
|
||||
.content
|
||||
.as_ref()
|
||||
.unwrap(),
|
||||
)?;
|
||||
|
||||
Ok(selected_llm)
|
||||
if let Some(ContentType::Text(content)) =
|
||||
&chat_completion_response.choices[0].message.content
|
||||
{
|
||||
info!(
|
||||
"router response: {}, response time: {}ms",
|
||||
content.replace("\n", "\\n"),
|
||||
router_response_time.as_millis()
|
||||
);
|
||||
let selected_llm = self.router_model.parse_response(content)?;
|
||||
Ok(selected_llm)
|
||||
} else {
|
||||
Ok(None)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
|
|||
|
|
@ -12,4 +12,5 @@ pub type Result<T> = std::result::Result<T, RoutingModelError>;
|
|||
pub trait RouterModel: Send + Sync {
|
||||
fn generate_request(&self, messages: &[Message]) -> ChatCompletionsRequest;
|
||||
fn parse_response(&self, content: &str) -> Result<Option<String>>;
|
||||
fn get_model_name(&self) -> String;
|
||||
}
|
||||
|
|
|
|||
|
|
@ -1,9 +1,8 @@
|
|||
use common::{
|
||||
api::open_ai::{ChatCompletionsRequest, Message},
|
||||
api::open_ai::{ChatCompletionsRequest, ContentType, Message},
|
||||
consts::{SYSTEM_ROLE, USER_ROLE},
|
||||
};
|
||||
use serde::{Deserialize, Serialize};
|
||||
use tracing::info;
|
||||
|
||||
use super::router_model::{RouterModel, RoutingModelError};
|
||||
|
||||
|
|
@ -68,7 +67,7 @@ impl RouterModel for RouterModelV1 {
|
|||
ChatCompletionsRequest {
|
||||
model: self.routing_model.clone(),
|
||||
messages: vec![Message {
|
||||
content: Some(message),
|
||||
content: Some(ContentType::Text(message)),
|
||||
role: USER_ROLE.to_string(),
|
||||
model: None,
|
||||
tool_calls: None,
|
||||
|
|
@ -86,10 +85,6 @@ impl RouterModel for RouterModelV1 {
|
|||
return Ok(None);
|
||||
}
|
||||
let router_resp_fixed = fix_json_response(content);
|
||||
info!(
|
||||
"router response (fixed): {}",
|
||||
router_resp_fixed.replace("\n", "\\n")
|
||||
);
|
||||
let router_response: LlmRouterResponse = serde_json::from_str(router_resp_fixed.as_str())?;
|
||||
|
||||
let selected_llm = router_response.route.unwrap_or_default().to_string();
|
||||
|
|
@ -100,6 +95,10 @@ impl RouterModel for RouterModelV1 {
|
|||
|
||||
Ok(Some(selected_llm))
|
||||
}
|
||||
|
||||
fn get_model_name(&self) -> String {
|
||||
self.routing_model.clone()
|
||||
}
|
||||
}
|
||||
|
||||
fn fix_json_response(body: &str) -> String {
|
||||
|
|
@ -172,22 +171,28 @@ user: "seattle"
|
|||
let messages = vec![
|
||||
Message {
|
||||
role: "system".to_string(),
|
||||
content: Some("You are a helpful assistant.".to_string()),
|
||||
content: Some(ContentType::Text(
|
||||
"You are a helpful assistant.".to_string(),
|
||||
)),
|
||||
..Default::default()
|
||||
},
|
||||
Message {
|
||||
role: "user".to_string(),
|
||||
content: Some("Hello, I want to book a flight.".to_string()),
|
||||
content: Some(ContentType::Text(
|
||||
"Hello, I want to book a flight.".to_string(),
|
||||
)),
|
||||
..Default::default()
|
||||
},
|
||||
Message {
|
||||
role: "assistant".to_string(),
|
||||
content: Some("Sure, where would you like to go?".to_string()),
|
||||
content: Some(ContentType::Text(
|
||||
"Sure, where would you like to go?".to_string(),
|
||||
)),
|
||||
..Default::default()
|
||||
},
|
||||
Message {
|
||||
role: "user".to_string(),
|
||||
content: Some("seattle".to_string()),
|
||||
content: Some(ContentType::Text("seattle".to_string())),
|
||||
..Default::default()
|
||||
},
|
||||
];
|
||||
|
|
@ -198,7 +203,7 @@ user: "seattle"
|
|||
|
||||
println!("Prompt: {}", prompt);
|
||||
|
||||
assert_eq!(expected_prompt, prompt);
|
||||
assert_eq!(expected_prompt, prompt.to_string());
|
||||
}
|
||||
}
|
||||
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue