lxy
e9e3578d26
- openai_compat: 扫描所有 assistant 消息剥离 orphan tool_calls（原仅查末条） - queue 加 conversationId 字段，按会话精准 drain - regenerate/editMessage 只清本会话排队消息 - newConversation 保留旧会话排队消息 - AiError 只清出错会话的队列项
724 lines
31 KiB
Rust
724 lines
31 KiB
Rust
//! OpenAI 兼容 Provider — 通过 /v1/chat/completions 端点实现
|
||
//!
|
||
//! 覆盖: OpenAI / GLM (open.bigmodel.cn) / DeepSeek / Claude OpenAI 兼容模式
|
||
//! 支持: 同步调用 + SSE 流式 + Function Calling / Tool Use
|
||
|
||
use std::time::Duration;
|
||
|
||
use async_trait::async_trait;
|
||
use futures::StreamExt;
|
||
use reqwest::Client;
|
||
use tracing::{debug, error, warn};
|
||
|
||
use crate::provider::{
|
||
CompletionRequest, CompletionResponse, LlmProvider, StreamResult, TokenUsage, ToolCall,
|
||
};
|
||
// ChatMessage 仅单测构造 CompletionRequest 用，避免非 test 构建的 unused import 警告。
|
||
#[cfg(test)]
|
||
use crate::provider::ChatMessage;
|
||
use crate::retry::{
|
||
retry_with_backoff, AttemptOutcome, is_reqwest_error_retryable, is_status_retryable,
|
||
};
|
||
|
||
// 协议数据结构（OpenAiRequest / OpenAiResponse / OpenAiStreamChunk …）与 SSE 解析纯函数
|
||
// apply_openai_sse 抽离到 openai_helpers.rs，结构对齐 anthropic_helpers.rs。
|
||
// Rust impl 块不可跨文件，Provider struct + impl 仍留本文件。
|
||
use crate::openai_helpers::{
|
||
apply_openai_sse, OpenAiMessage, OpenAiRequest, OpenAiResponse, OpenAiToolCallResp,
|
||
};
|
||
|
||
// ============================================================
|
||
// OpenAI Compat Provider
|
||
// ============================================================
|
||
|
||
/// OpenAI 兼容 LLM Provider
|
||
pub struct OpenAICompatProvider {
|
||
client: Client,
|
||
api_key: String,
|
||
base_url: String,
|
||
default_model: String,
|
||
}
|
||
|
||
impl OpenAICompatProvider {
|
||
/// 创建 Provider
|
||
///
|
||
/// - `base_url`: 如 "https://api.openai.com", "https://open.bigmodel.cn/api/paas", "https://api.deepseek.com"
|
||
/// - `api_key`: API 密钥
|
||
/// - `default_model`: 默认模型名称
|
||
pub fn new(base_url: impl Into<String>, api_key: impl Into<String>, default_model: impl Into<String>) -> Self {
|
||
// SW-260618-10: reqwest Client 构建抽 crate::build_provider_client（与 Anthropic 共用 DRY）。
|
||
// connect_timeout/回退策略集中此处，未来改一处即可（见 lib.rs::build_provider_client）。
|
||
let client = crate::build_provider_client();
|
||
Self {
|
||
client,
|
||
api_key: api_key.into(),
|
||
base_url: base_url.into(),
|
||
default_model: default_model.into(),
|
||
}
|
||
}
|
||
|
||
/// 构建完整 API URL
|
||
///
|
||
/// 智能拼接，兼容三种 base_url 约定：
|
||
/// - 已含完整端点（…/chat/completions）→ 直接用
|
||
/// - 已含版本段（…/v1 …/v4 等，如 GLM 的 /api/paas/v4）→ 补 /chat/completions
|
||
/// - 仅域名无版本（如 api.openai.com / api.deepseek.com）→ 补 /v1/chat/completions（OpenAI 约定）
|
||
fn chat_url(&self) -> String {
|
||
let base = self.base_url.trim_end_matches('/');
|
||
if base.ends_with("/chat/completions") {
|
||
return base.to_string();
|
||
}
|
||
if Self::ends_with_version(base) {
|
||
return format!("{}/chat/completions", base);
|
||
}
|
||
format!("{}/v1/chat/completions", base)
|
||
}
|
||
|
||
/// base_url 是否以 `/v<数字>` 结尾（如 /v1 /v4）
|
||
fn ends_with_version(base: &str) -> bool {
|
||
match base.rsplit_once('/') {
|
||
Some((_, last)) if last.starts_with('v') && last.len() > 1 => {
|
||
last[1..].bytes().all(|b| b.is_ascii_digit())
|
||
}
|
||
_ => false,
|
||
}
|
||
}
|
||
|
||
/// 构建 embeddings API URL（与 chat_url 同套智能拼接规则）
|
||
fn embed_url(&self) -> String {
|
||
let base = self.base_url.trim_end_matches('/');
|
||
if base.ends_with("/embeddings") {
|
||
return base.to_string();
|
||
}
|
||
if Self::ends_with_version(base) {
|
||
return format!("{}/embeddings", base);
|
||
}
|
||
format!("{}/v1/embeddings", base)
|
||
}
|
||
|
||
/// 将通用请求转换为 OpenAI 格式
|
||
fn convert_request(&self, req: CompletionRequest) -> OpenAiRequest {
|
||
let model = if req.model.is_empty() {
|
||
self.default_model.clone()
|
||
} else {
|
||
req.model
|
||
};
|
||
|
||
let mut messages: Vec<OpenAiMessage> = req
|
||
.messages
|
||
.into_iter()
|
||
.map(|m| {
|
||
let role = match m.role {
|
||
crate::provider::MessageRole::System => "system",
|
||
crate::provider::MessageRole::User => "user",
|
||
crate::provider::MessageRole::Assistant => "assistant",
|
||
crate::provider::MessageRole::Tool => "tool",
|
||
};
|
||
// F-260614-05 Phase 2a: 多模态 content（须在 move m.tool_calls 之前算，借用 m）。
|
||
// 含图消息走 content 数组（text/image_url）；纯文本走字符串简写
|
||
// （保持与现有纯文本端点零回归）。image_url 支持 data URI（base64）与 http(s) URL。
|
||
let content = if m.has_image() {
|
||
let parts: Vec<serde_json::Value> = m
|
||
.flattened_parts()
|
||
.into_iter()
|
||
.map(|p| match p {
|
||
crate::provider::ContentPart::Text { text } => serde_json::json!({
|
||
"type": "text",
|
||
"text": text,
|
||
}),
|
||
crate::provider::ContentPart::Image { url, base64, media_type, alt: _ } => {
|
||
let final_url = match (base64, url, media_type) {
|
||
(Some(b), _, Some(mt)) => {
|
||
format!("data:{};base64,{}", mt, b)
|
||
}
|
||
(None, Some(u), _) => u,
|
||
// 完整性兜底：当前 image_base64 构造器强制 media_type:Some，
|
||
// image_url 构造器提供 url:Some，二者分别命中上两个分支；
|
||
// 此分支仅在 parts 来源被外部直接构造且字段均缺时才可达
|
||
//（如 url:None+base64:None 或 base64:Some+media_type:None）。
|
||
// 退化为空串（OpenAI 对空 image_url.url 会 400），
|
||
// 由调用方保证 parts 合法性，provider 层不做静默伪造。
|
||
_ => String::new(),
|
||
};
|
||
serde_json::json!({
|
||
"type": "image_url",
|
||
"image_url": { "url": final_url },
|
||
})
|
||
}
|
||
})
|
||
.collect();
|
||
serde_json::Value::Array(parts)
|
||
} else {
|
||
serde_json::Value::String(m.content.clone())
|
||
};
|
||
let tool_calls = m.tool_calls.map(|calls| {
|
||
calls
|
||
.into_iter()
|
||
.map(|tc| {
|
||
serde_json::json!({
|
||
"id": tc.id,
|
||
"type": tc.call_type,
|
||
"function": {
|
||
"name": tc.function.name,
|
||
"arguments": tc.function.arguments,
|
||
}
|
||
})
|
||
})
|
||
.collect()
|
||
});
|
||
OpenAiMessage {
|
||
role: role.to_string(),
|
||
content,
|
||
tool_call_id: m.tool_call_id,
|
||
tool_calls,
|
||
reasoning_content: m.reasoning_content,
|
||
}
|
||
})
|
||
.collect();
|
||
|
||
// B-260626-01: 保证首条 user/system(OpenAI 协议要求首条非 assistant/tool)。
|
||
// 对齐 AnthropicCompatProvider::ensure_leading_user:上游绕过 sanitize 的调用方
|
||
// (标题生成/知识注入/工作流 AI 节点等直构造 CompletionRequest 的路径)可能传入首条
|
||
// assistant 的序列(会话恢复/续发/片段截取),补 user 占位保留上下文,首条合法。
|
||
Self::ensure_leading_user(&mut messages);
|
||
|
||
// 治 DeepSeek 400「insufficient tool messages」:扫描所有 assistant 消息,
|
||
// 若某条 assistant 含 tool_calls 但下一条不是 tool,则剥离其 tool_calls。
|
||
// 正常流程 tool 结果先于下一轮 LLM 请求推入历史,此守卫仅兜底异常截断/恢复场景的残末尾。
|
||
// 注意:合法的三元组形如:assistant(tc=[a]) → tool(a) → assistant(tc=[b]) → tool(b)。
|
||
// 若最后一条是 assistant(tc=...) 也无下一条 tool,同样剥离。
|
||
for i in 0..messages.len() {
|
||
let role = messages[i].role.clone();
|
||
if role != "assistant" {
|
||
continue;
|
||
}
|
||
let has_tc = messages[i].tool_calls.is_some();
|
||
if !has_tc {
|
||
continue;
|
||
}
|
||
let next_is_tool = i + 1 < messages.len()
|
||
&& matches!(messages[i + 1].role.as_str(), "tool");
|
||
if !next_is_tool {
|
||
messages[i].tool_calls = None;
|
||
tracing::warn!(
|
||
"[openai] assistant(#{} role={}) 含 tool_calls 但下一条非 tool,已自动剥离(防 400)",
|
||
i, role,
|
||
);
|
||
}
|
||
}
|
||
|
||
let tools = req.tools.map(|defs| {
|
||
defs.into_iter()
|
||
.map(|d| serde_json::to_value(d).unwrap_or_default())
|
||
.collect()
|
||
});
|
||
|
||
OpenAiRequest {
|
||
model,
|
||
messages,
|
||
temperature: req.temperature,
|
||
max_tokens: req.max_tokens,
|
||
stream: req.stream,
|
||
tools,
|
||
tool_choice: req.tool_choice,
|
||
reasoning_content: req.reasoning_content,
|
||
// 流式请求末 chunk 带 usage（同步调用 complete 不需要）
|
||
stream_options: if req.stream {
|
||
Some(serde_json::json!({ "include_usage": true }))
|
||
} else {
|
||
None
|
||
},
|
||
}
|
||
}
|
||
|
||
/// B-260626-01: 保证 messages 首条为 user/system(OpenAI 协议要求首条非 assistant/tool)。
|
||
///
|
||
/// 对齐 `AnthropicCompatProvider::ensure_leading_user`。上游绕过 `ContextManager::sanitize_messages`
|
||
/// 的调用方(标题生成/知识注入/工作流 AI 节点等直构造 CompletionRequest 的路径)可能传入首条
|
||
/// assistant 的序列——会话恢复、续发或历史片段截取时,真正的首条 user 已被裁剪/压缩掉。
|
||
///
|
||
/// **用"补"而非"砍"**:开头插一条 user 占位,保留全部上下文(砍会丢工具调用历史,多轮砍到空)。
|
||
/// 占位 user 紧贴原首条,不破坏 user/assistant 交替;仅异常路径触发(正常首条本就是 user)。
|
||
fn ensure_leading_user(messages: &mut Vec<OpenAiMessage>) {
|
||
let first_role = messages.first().map(|m| m.role.as_str()).unwrap_or("");
|
||
if first_role == "user" || first_role == "system" {
|
||
return;
|
||
}
|
||
warn!(
|
||
first_role,
|
||
msg_count = messages.len(),
|
||
"ensure_leading_user: 首条非 user/system,补 user 占位(保留上下文,防 OpenAI 首条 assistant/tool 非法)"
|
||
);
|
||
messages.insert(
|
||
0,
|
||
OpenAiMessage {
|
||
role: "user".into(),
|
||
content: serde_json::Value::String("(continued from previous context)".into()),
|
||
tool_call_id: None,
|
||
tool_calls: None,
|
||
reasoning_content: None,
|
||
},
|
||
);
|
||
}
|
||
|
||
/// 解析同步响应中的工具调用
|
||
fn parse_tool_calls(calls: Vec<OpenAiToolCallResp>) -> Vec<ToolCall> {
|
||
calls
|
||
.into_iter()
|
||
.map(|c| ToolCall::new(c.id, c.function.name, c.function.arguments))
|
||
.collect()
|
||
}
|
||
}
|
||
|
||
#[async_trait]
|
||
impl LlmProvider for OpenAICompatProvider {
|
||
/// 文本嵌入: POST /v1/embeddings(OpenAI 兼容,智谱/阿里百炼/OpenAI 通用)
|
||
async fn embed(&self, model: &str, texts: Vec<String>) -> anyhow::Result<Vec<Vec<f32>>> {
|
||
#[derive(serde::Deserialize)]
|
||
struct EmbedData { embedding: Vec<f32>, index: usize }
|
||
#[derive(serde::Deserialize)]
|
||
struct EmbedResponse { data: Vec<EmbedData> }
|
||
|
||
let resp = self
|
||
.client
|
||
.post(self.embed_url())
|
||
.header("Authorization", format!("Bearer {}", self.api_key))
|
||
.header("Content-Type", "application/json")
|
||
.json(&serde_json::json!({ "model": model, "input": texts }))
|
||
.send()
|
||
.await?;
|
||
|
||
if !resp.status().is_success() {
|
||
let status = resp.status();
|
||
let body = resp.text().await.unwrap_or_default();
|
||
anyhow::bail!("Embedding API 错误 {}: {}", status, body);
|
||
}
|
||
|
||
let mut body: EmbedResponse = resp.json().await?;
|
||
// 按 index 排序保证与输入顺序一致(API 不保证返回顺序)
|
||
body.data.sort_by_key(|d| d.index);
|
||
Ok(body.data.into_iter().map(|d| d.embedding).collect())
|
||
}
|
||
|
||
async fn complete(&self, request: CompletionRequest) -> anyhow::Result<CompletionResponse> {
|
||
let mut req = request;
|
||
req.stream = false;
|
||
let openai_req = self.convert_request(req);
|
||
|
||
debug!(model = %openai_req.model, "OpenAI 同步调用");
|
||
|
||
// 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);
|
||
}
|
||
// 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,
|
||
reasoning_content: choice.message.reasoning_content,
|
||
})
|
||
}
|
||
})
|
||
.await
|
||
}
|
||
|
||
async fn stream(&self, request: CompletionRequest) -> anyhow::Result<StreamResult> {
|
||
let mut req = request;
|
||
req.stream = true;
|
||
let openai_req = self.convert_request(req);
|
||
|
||
debug!(model = %openai_req.model, "OpenAI 流式调用");
|
||
|
||
// BUG-2026-07-07: send 阶段需 timeout 防 hang(同 Anthropic 路径)。
|
||
// 不能用 reqwest .timeout()(会砍流式 body),改用 tokio::time::timeout 包裹 send。
|
||
let send_future = self
|
||
.client
|
||
.post(self.chat_url())
|
||
.header("Authorization", format!("Bearer {}", self.api_key))
|
||
.header("Content-Type", "application/json")
|
||
.json(&openai_req)
|
||
.send();
|
||
let resp = match tokio::time::timeout(Duration::from_secs(60), send_future).await {
|
||
Ok(Ok(r)) => r,
|
||
Ok(Err(e)) => {
|
||
tracing::error!(error = %e, is_timeout = e.is_timeout(), "OpenAI 流式 send 失败");
|
||
return Err(e.into());
|
||
}
|
||
Err(_elapsed) => {
|
||
tracing::error!(url = %self.chat_url(), "OpenAI 流式 send 超时(60s 未返回响应头)");
|
||
anyhow::bail!("流式请求超时(60秒未收到 HTTP 响应，可能服务不可达或被防火墙拦截)");
|
||
}
|
||
};
|
||
|
||
if !resp.status().is_success() {
|
||
let status = resp.status();
|
||
let body = resp.text().await.unwrap_or_default();
|
||
error!(%status, %body, "LLM 流式 API 调用失败");
|
||
anyhow::bail!("LLM 流式 API 错误 {}: {}", status, body);
|
||
}
|
||
|
||
// BUG-2026-07-17 根治: 原生 SSE 解析器替代 eventsource-stream 库。
|
||
// eventsource-stream 在 Windows 上对 Deepseek 等响应报 "error decoding response body"
|
||
// (严格 UTF-8 + SSE 协议校验,跨 chunk 字符/不完整事件均报错且不可恢复)。
|
||
// 原生解析器:bytes 累积 + from_utf8_lossy 宽松处理 + \n\n 分隔,容错不中断流。
|
||
let mut last_usage: Option<TokenUsage> = None;
|
||
|
||
let sse = crate::sse_parser::SseStream::new(resp.bytes_stream());
|
||
let stream = sse.flat_map(move |result: Result<Vec<String>, String>| {
|
||
let mut chunks: Vec<anyhow::Result<crate::provider::StreamChunk>> = Vec::new();
|
||
match result {
|
||
Ok(events) => {
|
||
for data in events {
|
||
let chunk = apply_openai_sse(&data, &mut last_usage);
|
||
chunks.push(Ok(chunk));
|
||
}
|
||
}
|
||
Err(e) => {
|
||
let ctx = format!("SSE 流错误: {}", e);
|
||
error!("{}", ctx);
|
||
chunks.push(Err(anyhow::anyhow!("{}", ctx)));
|
||
}
|
||
}
|
||
futures::stream::iter(chunks)
|
||
});
|
||
|
||
Ok(Box::pin(stream))
|
||
}
|
||
|
||
fn name(&self) -> &str {
|
||
&self.default_model
|
||
}
|
||
|
||
fn endpoint(&self) -> String {
|
||
self.chat_url()
|
||
}
|
||
}
|
||
|
||
// ============================================================
|
||
// 单测（不发真实 HTTP，喂构造的 SSE data 字符串序列）
|
||
// ============================================================
|
||
|
||
#[cfg(test)]
|
||
mod tests {
|
||
use super::*;
|
||
|
||
/// 辅助：构造普通文本 delta chunk 的 SSE data
|
||
fn text_chunk(content: &str, finish_reason: Option<&str>) -> String {
|
||
let fr = match finish_reason {
|
||
Some(r) => format!(", \"finish_reason\": \"{}\"", r),
|
||
None => String::from(", \"finish_reason\": null"),
|
||
};
|
||
format!(
|
||
r#"{{"choices":[{{"delta":{{"content":"{}"}}{}}}]}}"#,
|
||
content, fr
|
||
)
|
||
}
|
||
|
||
/// 辅助：构造带 usage 的 chunk（choices 为空 → usage-only 末 chunk，对应 include_usage）
|
||
fn usage_only_chunk(prompt: u32, completion: u32) -> String {
|
||
format!(
|
||
r#"{{"choices":[],"usage":{{"prompt_tokens":{},"completion_tokens":{},"total_tokens":{}}}}}"#,
|
||
prompt,
|
||
completion,
|
||
prompt + completion
|
||
)
|
||
}
|
||
|
||
/// 辅助：构造既有 content 又带 usage 的末段 chunk（部分兼容端点会把 usage 挂到正常末 chunk 上）
|
||
fn text_chunk_with_usage(content: &str, finish_reason: &str, prompt: u32, completion: u32) -> String {
|
||
format!(
|
||
r#"{{"choices":[{{"delta":{{"content":"{}"}},"finish_reason":"{}"}}],"usage":{{"prompt_tokens":{},"completion_tokens":{},"total_tokens":{}}}}}"#,
|
||
content,
|
||
finish_reason,
|
||
prompt,
|
||
completion,
|
||
prompt + completion
|
||
)
|
||
}
|
||
|
||
/// 多 chunk 文本流后，末 chunk 携带 usage（include_usage 覆盖语义）
|
||
#[test]
|
||
fn openai_sse_multi_chunk_with_terminal_usage() {
|
||
let mut acc: Option<TokenUsage> = None;
|
||
|
||
// 1) 首个文本增量，无 usage
|
||
let c = apply_openai_sse(&text_chunk("Hello", None), &mut acc);
|
||
assert_eq!(c.delta, "Hello");
|
||
assert!(!c.finished);
|
||
assert!(c.usage.is_none());
|
||
assert!(acc.is_none(), "无 usage 的 chunk 不应改累加器");
|
||
|
||
// 2) 第二个文本增量
|
||
let c = apply_openai_sse(&text_chunk(" world", None), &mut acc);
|
||
assert_eq!(c.delta, " world");
|
||
assert!(!c.finished);
|
||
assert!(acc.is_none());
|
||
|
||
// 3) 末段正常 chunk 带 finish_reason=stop（仍是文本 delta，不带 usage）
|
||
let c = apply_openai_sse(&text_chunk("", Some("stop")), &mut acc);
|
||
assert!(c.finished);
|
||
assert_eq!(c.delta, "");
|
||
assert!(acc.is_none(), "此 chunk 无 usage 字段，累加器仍为 None");
|
||
|
||
// 4) usage-only chunk（choices=[]）携带累计 usage → 覆盖累加器
|
||
let c = apply_openai_sse(&usage_only_chunk(12, 34), &mut acc);
|
||
assert!(!c.finished);
|
||
assert!(c.usage.is_none(), "非 [DONE] chunk 不带出 usage");
|
||
let acc = acc.expect("累加器应已被 usage-only chunk 覆盖写入");
|
||
assert_eq!(acc.prompt_tokens, 12);
|
||
assert_eq!(acc.completion_tokens, 34);
|
||
assert_eq!(acc.total_tokens, 46);
|
||
}
|
||
|
||
/// usage 挂在正常末段 chunk（含 finish_reason）上，而非独立 usage-only chunk
|
||
#[test]
|
||
fn openai_sse_usage_on_terminal_text_chunk() {
|
||
let mut acc: Option<TokenUsage> = None;
|
||
let c = apply_openai_sse(&text_chunk_with_usage("", "stop", 100, 200), &mut acc);
|
||
assert!(c.finished);
|
||
assert!(c.usage.is_none(), "非 [DONE] 不带出 usage，仅覆盖累加器");
|
||
let acc = acc.expect("末段 chunk 的 usage 应已覆盖累加器");
|
||
assert_eq!(acc.prompt_tokens, 100);
|
||
assert_eq!(acc.completion_tokens, 200);
|
||
assert_eq!(acc.total_tokens, 300);
|
||
}
|
||
|
||
/// [DONE] 时 take() 带出累积 usage，且取走后累加器清空
|
||
#[test]
|
||
fn openai_sse_done_takes_accumulated_usage() {
|
||
let mut acc: Option<TokenUsage> = None;
|
||
apply_openai_sse(&text_chunk("x", None), &mut acc);
|
||
apply_openai_sse(&usage_only_chunk(5, 7), &mut acc);
|
||
|
||
let c = apply_openai_sse("[DONE]", &mut acc);
|
||
assert!(c.finished);
|
||
let u = c.usage.expect("[DONE] 应带出累积 usage");
|
||
assert_eq!(u.prompt_tokens, 5);
|
||
assert_eq!(u.completion_tokens, 7);
|
||
assert_eq!(u.total_tokens, 12);
|
||
assert!(acc.is_none(), "take() 后累加器应清空");
|
||
}
|
||
|
||
/// 无 usage 的流：[DONE] 时 usage 字段为 None
|
||
#[test]
|
||
fn openai_sse_done_without_usage() {
|
||
let mut acc: Option<TokenUsage> = None;
|
||
apply_openai_sse(&text_chunk("hi", None), &mut acc);
|
||
let c = apply_openai_sse("[DONE]", &mut acc);
|
||
assert!(c.finished);
|
||
assert!(c.usage.is_none(), "全程无 usage 时 [DONE] usage 应为 None");
|
||
assert!(acc.is_none());
|
||
}
|
||
|
||
/// 后续 usage chunk 覆盖先前 usage（多轮 / 重发场景）
|
||
#[test]
|
||
fn openai_sse_later_usage_overrides_earlier() {
|
||
let mut acc: Option<TokenUsage> = None;
|
||
apply_openai_sse(&usage_only_chunk(1, 1), &mut acc);
|
||
apply_openai_sse(&usage_only_chunk(50, 60), &mut acc);
|
||
let c = apply_openai_sse("[DONE]", &mut acc);
|
||
let u = c.usage.unwrap();
|
||
assert_eq!(u.prompt_tokens, 50, "末 usage 应覆盖前值");
|
||
assert_eq!(u.completion_tokens, 60);
|
||
assert_eq!(u.total_tokens, 110);
|
||
}
|
||
|
||
/// finish_reason=length（max_tokens 截断）按正常终止处理
|
||
#[test]
|
||
fn openai_sse_length_finish_reason_treated_as_finished() {
|
||
let mut acc: Option<TokenUsage> = None;
|
||
let c = apply_openai_sse(&text_chunk("...", Some("length")), &mut acc);
|
||
assert!(c.finished, "length 应视为正常终止");
|
||
assert!(acc.is_none());
|
||
}
|
||
|
||
/// 非法 JSON data → 返回空 chunk，不 panic、不改累加器
|
||
#[test]
|
||
fn openai_sse_malformed_json_yields_empty_chunk() {
|
||
let mut acc: Option<TokenUsage> = None;
|
||
let c = apply_openai_sse("not a json", &mut acc);
|
||
assert_eq!(c.delta, "");
|
||
assert!(!c.finished);
|
||
assert!(c.usage.is_none());
|
||
assert!(acc.is_none());
|
||
}
|
||
|
||
/// tool_calls 增量解析
|
||
#[test]
|
||
fn openai_sse_tool_call_delta() {
|
||
let mut acc: Option<TokenUsage> = None;
|
||
let data = r#"{"choices":[{"delta":{"tool_calls":[{"index":0,"id":"call_1","function":{"name":"get_weather","arguments":"{\"q\":"}}]}}]}"#;
|
||
let c = apply_openai_sse(data, &mut acc);
|
||
assert!(acc.is_none());
|
||
let tcs = c.tool_calls.expect("应有 tool_calls 增量");
|
||
assert_eq!(tcs.len(), 1);
|
||
assert_eq!(tcs[0].index, 0);
|
||
assert_eq!(tcs[0].id.as_deref(), Some("call_1"));
|
||
assert_eq!(tcs[0].function_name.as_deref(), Some("get_weather"));
|
||
assert_eq!(tcs[0].function_arguments.as_deref(), Some("{\"q\":"));
|
||
assert!(!c.finished);
|
||
}
|
||
|
||
// ---------- F-260614-05 Phase 2a 多模态 convert_request ----------
|
||
|
||
/// 含图消息 → content 数组（text + image_url data URI）；纯文本 → 字符串简写
|
||
#[test]
|
||
fn openai_convert_multimodal_content() {
|
||
let provider = OpenAICompatProvider::new("https://api.openai.com", "k", "gpt-4o");
|
||
let req = CompletionRequest {
|
||
model: "gpt-4o".into(),
|
||
messages: vec![ChatMessage::user_parts(
|
||
"看图",
|
||
vec![
|
||
crate::provider::ContentPart::image_base64("image/png", "iVBOR"),
|
||
crate::provider::ContentPart::image_url("https://x/a.png"),
|
||
],
|
||
)],
|
||
temperature: None,
|
||
max_tokens: None,
|
||
stream: false,
|
||
tools: None,
|
||
tool_choice: None,
|
||
reasoning_content: None,
|
||
};
|
||
let out = provider.convert_request(req);
|
||
let msg = &out.messages[0];
|
||
// content 是数组：[text "看图", image_url(data URI), image_url(http url)]
|
||
let arr = msg.content.as_array().expect("含图 → content 数组");
|
||
assert_eq!(arr.len(), 3);
|
||
assert_eq!(arr[0]["type"], "text");
|
||
assert_eq!(arr[0]["text"], "看图");
|
||
assert_eq!(arr[1]["type"], "image_url");
|
||
assert_eq!(
|
||
arr[1]["image_url"]["url"],
|
||
"data:image/png;base64,iVBOR"
|
||
);
|
||
assert_eq!(arr[2]["image_url"]["url"], "https://x/a.png");
|
||
}
|
||
|
||
/// 纯文本消息 → content 仍是字符串简写（无图不数组化，对齐纯文本端点兼容）
|
||
#[test]
|
||
fn openai_convert_text_only_remains_string() {
|
||
let provider = OpenAICompatProvider::new("https://api.openai.com", "k", "gpt-4o");
|
||
let req = CompletionRequest {
|
||
model: "gpt-4o".into(),
|
||
messages: vec![ChatMessage::user("hello")],
|
||
temperature: None,
|
||
max_tokens: None,
|
||
stream: false,
|
||
tools: None,
|
||
tool_choice: None,
|
||
reasoning_content: None,
|
||
};
|
||
let out = provider.convert_request(req);
|
||
let msg = &out.messages[0];
|
||
assert_eq!(msg.content, serde_json::Value::String("hello".into()));
|
||
}
|
||
|
||
// ---------- B-260626-01: ensure_leading_user（首条非 user/system → 补 user 占位,OpenAI 对称 Anthropic）----------
|
||
|
||
/// B-260626-01: 首条 assistant → 补 user 占位(对齐 Anthropic)。上游绕过 sanitize 的
|
||
/// 调用方(title/knowledge_inject/工作流节点)可能传入首条 assistant 序列,补占位保留上下文。
|
||
#[test]
|
||
fn openai_ensure_leading_user_first_assistant_gets_placeholder() {
|
||
let provider = OpenAICompatProvider::new("https://api.openai.com", "k", "gpt-4o");
|
||
let req = CompletionRequest {
|
||
model: "gpt-4o".into(),
|
||
messages: vec![
|
||
ChatMessage::assistant("我来帮你"),
|
||
ChatMessage::user("继续"),
|
||
],
|
||
temperature: None,
|
||
max_tokens: None,
|
||
stream: false,
|
||
tools: None,
|
||
tool_choice: None,
|
||
reasoning_content: None,
|
||
};
|
||
let out = provider.convert_request(req);
|
||
assert_eq!(out.messages.len(), 3, "占位 + 原 2 条");
|
||
assert_eq!(out.messages[0].role.as_str(), "user", "首条应为 user(补占位)");
|
||
assert_eq!(out.messages[1].role.as_str(), "assistant");
|
||
assert_eq!(out.messages[2].role.as_str(), "user");
|
||
}
|
||
|
||
/// B-260626-01: 正常序列(user 开头)不补占位——零回归。
|
||
#[test]
|
||
fn openai_ensure_leading_user_normal_unchanged() {
|
||
let provider = OpenAICompatProvider::new("https://api.openai.com", "k", "gpt-4o");
|
||
let req = CompletionRequest {
|
||
model: "gpt-4o".into(),
|
||
messages: vec![
|
||
ChatMessage::user("hello"),
|
||
ChatMessage::assistant("hi"),
|
||
],
|
||
temperature: None,
|
||
max_tokens: None,
|
||
stream: false,
|
||
tools: None,
|
||
tool_choice: None,
|
||
reasoning_content: None,
|
||
};
|
||
let out = provider.convert_request(req);
|
||
assert_eq!(out.messages.len(), 2, "正常序列不补占位");
|
||
assert_eq!(out.messages[0].role.as_str(), "user");
|
||
}
|
||
}
|