优化: AI模块(provider重试兼容+aichat交互+工具卡片可读化+diff高亮)
This commit is contained in:
@@ -16,13 +16,16 @@ use crate::provider::{
|
||||
CompletionRequest, CompletionResponse, LlmProvider, StreamChunk, StreamResult,
|
||||
TokenUsage, ToolCall, ToolCallDelta,
|
||||
};
|
||||
use crate::retry::{
|
||||
retry_with_backoff, AttemptOutcome, is_reqwest_error_retryable, is_status_retryable,
|
||||
};
|
||||
|
||||
// ============================================================
|
||||
// OpenAI API 请求/响应结构体
|
||||
// ============================================================
|
||||
|
||||
/// OpenAI 兼容请求体
|
||||
#[derive(Debug, Serialize)]
|
||||
#[derive(Debug, Clone, Serialize)]
|
||||
struct OpenAiRequest {
|
||||
model: String,
|
||||
messages: Vec<OpenAiMessage>,
|
||||
@@ -41,7 +44,7 @@ struct OpenAiRequest {
|
||||
}
|
||||
|
||||
/// OpenAI 消息格式
|
||||
#[derive(Debug, Serialize, Deserialize)]
|
||||
#[derive(Debug, Clone, Serialize, Deserialize)]
|
||||
struct OpenAiMessage {
|
||||
role: String,
|
||||
content: String,
|
||||
@@ -398,60 +401,75 @@ impl LlmProvider for OpenAICompatProvider {
|
||||
|
||||
debug!(model = %openai_req.model, "OpenAI 同步调用");
|
||||
|
||||
// 同步路径整体超时(FR-R4)：client 仅配 connect_timeout，无总 timeout（防误砍流式长生成）。
|
||||
// complete 为一次性请求，挂死会让整轮对话静默卡住，此处用 RequestBuilder 的单请求 timeout 兜底，
|
||||
// 超时返回 is_timeout 错误（不静默挂），且不影响 stream() 流式路径。
|
||||
let resp = self
|
||||
.client
|
||||
.post(self.chat_url())
|
||||
.header("Authorization", format!("Bearer {}", self.api_key))
|
||||
.header("Content-Type", "application/json")
|
||||
.timeout(Duration::from_secs(60))
|
||||
.json(&openai_req)
|
||||
.send()
|
||||
.await
|
||||
.map_err(|e| {
|
||||
if e.is_timeout() {
|
||||
error!(model = %openai_req.model, "OpenAI 同步调用超时(60s)");
|
||||
anyhow::anyhow!("LLM 同步调用超时(60s)，请检查网络或模型可用性: {}", e)
|
||||
} else {
|
||||
anyhow::anyhow!(e)
|
||||
// 指数退避重试(B-260616-07): 包裹 send + 状态码判定。
|
||||
// 单请求 60s timeout 保持不变(FR-R4)，重试是额外层: 3 次 × 60s 最坏 180s，
|
||||
// 由 retry_with_backoff 内部 30s 总预算主动止损。
|
||||
let label = format!("OpenAI[{}]", openai_req.model);
|
||||
retry_with_backoff(&label, move |_| {
|
||||
let client = self.client.clone();
|
||||
let url = self.chat_url();
|
||||
let api_key = self.api_key.clone();
|
||||
let openai_req = openai_req.clone();
|
||||
async move {
|
||||
// send
|
||||
let resp = client
|
||||
.post(url)
|
||||
.header("Authorization", format!("Bearer {}", api_key))
|
||||
.header("Content-Type", "application/json")
|
||||
.timeout(Duration::from_secs(60))
|
||||
.json(&openai_req)
|
||||
.send()
|
||||
.await;
|
||||
let resp = match resp {
|
||||
Ok(r) => r,
|
||||
Err(e) => {
|
||||
if is_reqwest_error_retryable(&e) {
|
||||
return AttemptOutcome::Retryable(format!("请求失败(可重试): {}", e));
|
||||
}
|
||||
return AttemptOutcome::Fatal(format!("请求失败(不可重试): {}", e));
|
||||
}
|
||||
};
|
||||
// 状态码判定
|
||||
if !resp.status().is_success() {
|
||||
let status = resp.status().as_u16();
|
||||
let body = resp.text().await.unwrap_or_default();
|
||||
let msg = format!("LLM API 错误 {}: {}", status, body);
|
||||
if is_status_retryable(status) {
|
||||
warn!(%status, "OpenAI 同步调用可重试状态码");
|
||||
return AttemptOutcome::Retryable(msg);
|
||||
}
|
||||
error!(%status, %body, "LLM API 调用失败(不可重试)");
|
||||
return AttemptOutcome::Fatal(msg);
|
||||
}
|
||||
})?;
|
||||
|
||||
if !resp.status().is_success() {
|
||||
let status = resp.status();
|
||||
let body = resp.text().await.unwrap_or_default();
|
||||
error!(%status, %body, "LLM API 调用失败");
|
||||
anyhow::bail!("LLM API 错误 {}: {}", status, body);
|
||||
}
|
||||
|
||||
let body: OpenAiResponse = resp.json().await?;
|
||||
let choice = body
|
||||
.choices
|
||||
.into_iter()
|
||||
.next()
|
||||
.ok_or_else(|| anyhow::anyhow!("LLM 响应无 choices"))?;
|
||||
|
||||
let text = choice.message.content.unwrap_or_default();
|
||||
let tool_calls = choice.message.tool_calls.map(Self::parse_tool_calls);
|
||||
|
||||
let usage = body.usage.map(|u| TokenUsage {
|
||||
prompt_tokens: u.prompt_tokens,
|
||||
completion_tokens: u.completion_tokens,
|
||||
total_tokens: u.total_tokens,
|
||||
}).unwrap_or(TokenUsage {
|
||||
prompt_tokens: 0,
|
||||
completion_tokens: 0,
|
||||
total_tokens: 0,
|
||||
});
|
||||
|
||||
Ok(CompletionResponse {
|
||||
text,
|
||||
model: body.model,
|
||||
usage,
|
||||
tool_calls,
|
||||
// body 解析: 解析错属 Fatal(响应已成功送达，重试也会因同样格式失败)
|
||||
let body: OpenAiResponse = match resp.json().await {
|
||||
Ok(b) => b,
|
||||
Err(e) => return AttemptOutcome::Fatal(format!("响应解析失败: {}", e)),
|
||||
};
|
||||
let choice = match body.choices.into_iter().next() {
|
||||
Some(c) => c,
|
||||
None => return AttemptOutcome::Fatal("LLM 响应无 choices".to_string()),
|
||||
};
|
||||
let text = choice.message.content.unwrap_or_default();
|
||||
let tool_calls = choice.message.tool_calls.map(Self::parse_tool_calls);
|
||||
let usage = body.usage.map(|u| TokenUsage {
|
||||
prompt_tokens: u.prompt_tokens,
|
||||
completion_tokens: u.completion_tokens,
|
||||
total_tokens: u.total_tokens,
|
||||
}).unwrap_or(TokenUsage {
|
||||
prompt_tokens: 0,
|
||||
completion_tokens: 0,
|
||||
total_tokens: 0,
|
||||
});
|
||||
AttemptOutcome::Ok(CompletionResponse {
|
||||
text,
|
||||
model: body.model,
|
||||
usage,
|
||||
tool_calls,
|
||||
})
|
||||
}
|
||||
})
|
||||
.await
|
||||
}
|
||||
|
||||
async fn stream(&self, request: CompletionRequest) -> anyhow::Result<StreamResult> {
|
||||
|
||||
Reference in New Issue
Block a user