Files
GopherGate/src/providers/openai.rs
T
hobokenchicken e123f542f1
CI / Check (push) Has been cancelled
CI / Clippy (push) Has been cancelled
CI / Formatting (push) Has been cancelled
CI / Test (push) Has been cancelled
CI / Release Build (push) Has been cancelled
fix(openai): use max_output_tokens for Responses API
2026-03-17 18:36:05 +00:00

547 lines
24 KiB
Rust

use anyhow::Result;
use async_trait::async_trait;
use futures::stream::BoxStream;
use futures::StreamExt;
use super::helpers;
use super::{ProviderResponse, ProviderStreamChunk};
use crate::{config::AppConfig, errors::AppError, models::UnifiedRequest};
pub struct OpenAIProvider {
client: reqwest::Client,
config: crate::config::OpenAIConfig,
api_key: String,
pricing: Vec<crate::config::ModelPricing>,
}
impl OpenAIProvider {
pub fn new(config: &crate::config::OpenAIConfig, app_config: &AppConfig) -> Result<Self> {
let api_key = app_config.get_api_key("openai")?;
Self::new_with_key(config, app_config, api_key)
}
pub fn new_with_key(config: &crate::config::OpenAIConfig, app_config: &AppConfig, api_key: String) -> Result<Self> {
let client = reqwest::Client::builder()
.connect_timeout(std::time::Duration::from_secs(5))
.timeout(std::time::Duration::from_secs(300))
.pool_idle_timeout(std::time::Duration::from_secs(90))
.pool_max_idle_per_host(4)
.tcp_keepalive(std::time::Duration::from_secs(30))
.build()?;
Ok(Self {
client,
config: config.clone(),
api_key,
pricing: app_config.pricing.openai.clone(),
})
}
}
#[async_trait]
impl super::Provider for OpenAIProvider {
fn name(&self) -> &str {
"openai"
}
fn supports_model(&self, model: &str) -> bool {
model.starts_with("gpt-") ||
model.starts_with("o1-") ||
model.starts_with("o2-") ||
model.starts_with("o3-") ||
model.starts_with("o4-") ||
model.starts_with("o5-") ||
model.contains("gpt-5")
}
fn supports_multimodal(&self) -> bool {
true
}
async fn chat_completion(&self, request: UnifiedRequest) -> Result<ProviderResponse, AppError> {
// Allow proactive routing to Responses API based on heuristic
let model_lc = request.model.to_lowercase();
if model_lc.contains("gpt-5") || model_lc.contains("codex") {
return self.chat_responses(request).await;
}
let messages_json = helpers::messages_to_openai_json(&request.messages).await?;
let mut body = helpers::build_openai_body(&request, messages_json, false);
// Transition: Newer OpenAI models (o1, o3, gpt-5) require max_completion_tokens
// instead of the legacy max_tokens parameter.
if request.model.starts_with("o1-") || request.model.starts_with("o3-") || request.model.contains("gpt-5") {
if let Some(max_tokens) = body.as_object_mut().and_then(|obj| obj.remove("max_tokens")) {
body["max_completion_tokens"] = max_tokens;
}
}
let response = self
.client
.post(format!("{}/chat/completions", self.config.base_url))
.header("Authorization", format!("Bearer {}", self.api_key))
.json(&body)
.send()
.await
.map_err(|e| AppError::ProviderError(e.to_string()))?;
if !response.status().is_success() {
let status = response.status();
let error_text = response.text().await.unwrap_or_default();
// Read error body to diagnose. If the model requires the Responses
// API (v1/responses), retry against that endpoint.
if error_text.to_lowercase().contains("v1/responses") || error_text.to_lowercase().contains("only supported in v1/responses") {
return self.chat_responses(request).await;
}
tracing::error!("OpenAI API error ({}): {}", status, error_text);
return Err(AppError::ProviderError(format!("OpenAI API error ({}): {}", status, error_text)));
}
let resp_json: serde_json::Value = response
.json()
.await
.map_err(|e| AppError::ProviderError(e.to_string()))?;
helpers::parse_openai_response(&resp_json, request.model)
}
async fn chat_responses(&self, request: UnifiedRequest) -> Result<ProviderResponse, AppError> {
// Build a structured input for the Responses API.
let messages_json = helpers::messages_to_openai_json(&request.messages).await?;
let mut input_parts = Vec::new();
for m in &messages_json {
let role = m["role"].as_str().unwrap_or("user");
let mut content = m.get("content").cloned().unwrap_or(serde_json::json!([]));
// Map content types based on role for Responses API
if let Some(content_array) = content.as_array_mut() {
for part in content_array {
if let Some(part_obj) = part.as_object_mut() {
if let Some(t) = part_obj.get("type").and_then(|v| v.as_str()) {
match t {
"text" => {
let new_type = if role == "assistant" { "output_text" } else { "input_text" };
part_obj.insert("type".to_string(), serde_json::json!(new_type));
}
"image_url" => {
// Assistant typically doesn't have image_url in history this way, but for safety:
let new_type = if role == "assistant" { "output_image" } else { "input_image" };
part_obj.insert("type".to_string(), serde_json::json!(new_type));
if let Some(img_url) = part_obj.remove("image_url") {
part_obj.insert("image".to_string(), img_url);
}
}
_ => {}
}
}
}
}
} else if let Some(text) = content.as_str() {
let new_type = if role == "assistant" { "output_text" } else { "input_text" };
content = serde_json::json!([{ "type": new_type, "text": text }]);
}
input_parts.push(serde_json::json!({
"role": role,
"content": content
}));
}
let mut body = serde_json::json!({
"model": request.model,
"input": input_parts,
});
// Add standard parameters
if let Some(temp) = request.temperature {
body["temperature"] = serde_json::json!(temp);
}
// Newer models (gpt-5, o1) in Responses API use max_output_tokens
if let Some(max_tokens) = request.max_tokens {
if request.model.contains("gpt-5") || request.model.starts_with("o1-") || request.model.starts_with("o3-") {
body["max_output_tokens"] = serde_json::json!(max_tokens);
} else {
body["max_tokens"] = serde_json::json!(max_tokens);
}
}
if let Some(tools) = &request.tools {
body["tools"] = serde_json::json!(tools);
}
let resp = self
.client
.post(format!("{}/responses", self.config.base_url))
.header("Authorization", format!("Bearer {}", self.api_key))
.json(&body)
.send()
.await
.map_err(|e| AppError::ProviderError(e.to_string()))?;
if !resp.status().is_success() {
let err = resp.text().await.unwrap_or_default();
return Err(AppError::ProviderError(format!("OpenAI Responses API error: {}", err)));
}
let resp_json: serde_json::Value = resp.json().await.map_err(|e| AppError::ProviderError(e.to_string()))?;
// Try to normalize: if it's chat-style, use existing parser
if resp_json.get("choices").is_some() {
return helpers::parse_openai_response(&resp_json, request.model);
}
// Normalize Responses API output into ProviderResponse
let mut content_text = String::new();
if let Some(output) = resp_json.get("output").and_then(|o| o.as_array()) {
for out in output {
if let Some(contents) = out.get("content").and_then(|c| c.as_array()) {
for item in contents {
if let Some(text) = item.get("text").and_then(|t| t.as_str()) {
if !content_text.is_empty() { content_text.push_str("\n"); }
content_text.push_str(text);
} else if let Some(parts) = item.get("parts").and_then(|p| p.as_array()) {
for p in parts {
if let Some(t) = p.as_str() {
if !content_text.is_empty() { content_text.push_str("\n"); }
content_text.push_str(t);
}
}
}
}
}
}
}
if content_text.is_empty() {
if let Some(cands) = resp_json.get("candidates").and_then(|c| c.as_array()) {
if let Some(c0) = cands.get(0) {
if let Some(content) = c0.get("content") {
if let Some(parts) = content.get("parts").and_then(|p| p.as_array()) {
for p in parts {
if let Some(t) = p.get("text").and_then(|v| v.as_str()) {
if !content_text.is_empty() { content_text.push_str("\n"); }
content_text.push_str(t);
}
}
}
}
}
}
}
let prompt_tokens = resp_json.get("usage").and_then(|u| u.get("prompt_tokens")).and_then(|v| v.as_u64()).unwrap_or(0) as u32;
let completion_tokens = resp_json.get("usage").and_then(|u| u.get("completion_tokens")).and_then(|v| v.as_u64()).unwrap_or(0) as u32;
let total_tokens = resp_json.get("usage").and_then(|u| u.get("total_tokens")).and_then(|v| v.as_u64()).unwrap_or(0) as u32;
Ok(ProviderResponse {
content: content_text,
reasoning_content: None,
tool_calls: None,
prompt_tokens,
completion_tokens,
reasoning_tokens: 0,
total_tokens,
cache_read_tokens: 0,
cache_write_tokens: 0,
model: request.model,
})
}
fn estimate_tokens(&self, request: &UnifiedRequest) -> Result<u32> {
Ok(crate::utils::tokens::estimate_request_tokens(&request.model, request))
}
fn calculate_cost(
&self,
model: &str,
prompt_tokens: u32,
completion_tokens: u32,
cache_read_tokens: u32,
cache_write_tokens: u32,
registry: &crate::models::registry::ModelRegistry,
) -> f64 {
helpers::calculate_cost_with_registry(
model,
prompt_tokens,
completion_tokens,
cache_read_tokens,
cache_write_tokens,
registry,
&self.pricing,
0.15,
0.60,
)
}
async fn chat_completion_stream(
&self,
request: UnifiedRequest,
) -> Result<BoxStream<'static, Result<ProviderStreamChunk, AppError>>, AppError> {
// Allow proactive routing to Responses API based on heuristic
let model_lc = request.model.to_lowercase();
if model_lc.contains("gpt-5") || model_lc.contains("codex") {
return self.chat_responses_stream(request).await;
}
let messages_json = helpers::messages_to_openai_json(&request.messages).await?;
let mut body = helpers::build_openai_body(&request, messages_json, true);
// Standard OpenAI cleanup
if let Some(obj) = body.as_object_mut() {
// stream_options.include_usage is supported by OpenAI for token usage in streaming
// Transition: Newer OpenAI models (o1, o3, gpt-5) require max_completion_tokens
if request.model.starts_with("o1-") || request.model.starts_with("o3-") || request.model.contains("gpt-5") {
if let Some(max_tokens) = obj.remove("max_tokens") {
obj.insert("max_completion_tokens".to_string(), max_tokens);
}
}
}
let url = format!("{}/chat/completions", self.config.base_url);
let api_key = self.api_key.clone();
let probe_client = self.client.clone();
let probe_body = body.clone();
let model = request.model.clone();
let es = reqwest_eventsource::EventSource::new(
self.client
.post(&url)
.header("Authorization", format!("Bearer {}", self.api_key))
.json(&body),
)
.map_err(|e| AppError::ProviderError(format!("Failed to create EventSource: {}", e)))?;
let stream = async_stream::try_stream! {
let mut es = es;
while let Some(event) = es.next().await {
match event {
Ok(reqwest_eventsource::Event::Message(msg)) => {
if msg.data == "[DONE]" {
break;
}
let chunk: serde_json::Value = serde_json::from_str(&msg.data)
.map_err(|e| AppError::ProviderError(format!("Failed to parse stream chunk: {}", e)))?;
if let Some(p_chunk) = helpers::parse_openai_stream_chunk(&chunk, &model, None) {
yield p_chunk?;
}
}
Ok(_) => continue,
Err(e) => {
// Attempt to probe for the actual error body
let probe_resp = probe_client
.post(&url)
.header("Authorization", format!("Bearer {}", api_key))
.json(&probe_body)
.send()
.await;
match probe_resp {
Ok(r) if !r.status().is_success() => {
let status = r.status();
let error_body = r.text().await.unwrap_or_default();
tracing::error!("OpenAI Stream Error Probe ({}): {}", status, error_body);
tracing::debug!("Offending OpenAI Request Body: {}", serde_json::to_string(&probe_body).unwrap_or_default());
Err(AppError::ProviderError(format!("OpenAI API error ({}): {}", status, error_body)))?;
}
Ok(_) => {
// Probe returned success? This is unexpected if the original stream failed.
Err(AppError::ProviderError(format!("Stream error (probe returned 200): {}", e)))?;
}
Err(probe_err) => {
// Probe itself failed
tracing::error!("OpenAI Stream Error Probe failed: {}", probe_err);
Err(AppError::ProviderError(format!("Stream error (probe failed: {}): {}", probe_err, e)))?;
}
}
}
}
}
};
Ok(Box::pin(stream))
}
async fn chat_responses_stream(
&self,
request: UnifiedRequest,
) -> Result<BoxStream<'static, Result<ProviderStreamChunk, AppError>>, AppError> {
// Build a structured input for the Responses API.
let messages_json = helpers::messages_to_openai_json(&request.messages).await?;
let mut input_parts = Vec::new();
for m in &messages_json {
let role = m["role"].as_str().unwrap_or("user");
let mut content = m.get("content").cloned().unwrap_or(serde_json::json!([]));
// Map content types based on role for Responses API
if let Some(content_array) = content.as_array_mut() {
for part in content_array {
if let Some(part_obj) = part.as_object_mut() {
if let Some(t) = part_obj.get("type").and_then(|v| v.as_str()) {
match t {
"text" => {
let new_type = if role == "assistant" { "output_text" } else { "input_text" };
part_obj.insert("type".to_string(), serde_json::json!(new_type));
}
"image_url" => {
// Assistant typically doesn't have image_url in history this way, but for safety:
let new_type = if role == "assistant" { "output_image" } else { "input_image" };
part_obj.insert("type".to_string(), serde_json::json!(new_type));
if let Some(img_url) = part_obj.remove("image_url") {
part_obj.insert("image".to_string(), img_url);
}
}
_ => {}
}
}
}
}
} else if let Some(text) = content.as_str() {
let new_type = if role == "assistant" { "output_text" } else { "input_text" };
content = serde_json::json!([{ "type": new_type, "text": text }]);
}
input_parts.push(serde_json::json!({
"role": role,
"content": content
}));
}
let mut body = serde_json::json!({
"model": request.model,
"input": input_parts,
"stream": true,
"stream_options": { "include_usage": true }
});
// Add standard parameters
if let Some(temp) = request.temperature {
body["temperature"] = serde_json::json!(temp);
}
// Newer models (gpt-5, o1) in Responses API use max_output_tokens
if let Some(max_tokens) = request.max_tokens {
if request.model.contains("gpt-5") || request.model.starts_with("o1-") || request.model.starts_with("o3-") {
body["max_output_tokens"] = serde_json::json!(max_tokens);
} else {
body["max_tokens"] = serde_json::json!(max_tokens);
}
}
let url = format!("{}/responses", self.config.base_url);
let api_key = self.api_key.clone();
let model = request.model.clone();
let probe_client = self.client.clone();
let probe_body = body.clone();
let es = reqwest_eventsource::EventSource::new(
self.client
.post(&url)
.header("Authorization", format!("Bearer {}", api_key))
.json(&body),
)
.map_err(|e| AppError::ProviderError(format!("Failed to create EventSource for Responses API: {}", e)))?;
let stream = async_stream::try_stream! {
let mut es = es;
while let Some(event) = es.next().await {
match event {
Ok(reqwest_eventsource::Event::Message(msg)) => {
if msg.data == "[DONE]" {
break;
}
let chunk: serde_json::Value = serde_json::from_str(&msg.data)
.map_err(|e| AppError::ProviderError(format!("Failed to parse Responses stream chunk: {}", e)))?;
// Try standard OpenAI parsing first (choices/usage)
if let Some(p_chunk) = helpers::parse_openai_stream_chunk(&chunk, &model, None) {
yield p_chunk?;
} else {
// Responses API specific parsing for streaming
let mut content = String::new();
// Check for output[0].content[0].text (similar to non-stream)
if let Some(output) = chunk.get("output").and_then(|o| o.as_array()) {
for out in output {
if let Some(contents) = out.get("content").and_then(|c| c.as_array()) {
for item in contents {
if let Some(text) = item.get("text").and_then(|t| t.as_str()) {
content.push_str(text);
} else if let Some(delta) = item.get("delta").and_then(|d| d.get("text")).and_then(|t| t.as_str()) {
content.push_str(delta);
}
}
}
}
}
// Check for candidates[0].content.parts[0].text
if content.is_empty() {
if let Some(cands) = chunk.get("candidates").and_then(|c| c.as_array()) {
for c in cands {
if let Some(content_obj) = c.get("content") {
if let Some(parts) = content_obj.get("parts").and_then(|p| p.as_array()) {
for p in parts {
if let Some(t) = p.get("text").and_then(|v| v.as_str()) {
content.push_str(t);
}
}
}
}
}
}
}
if !content.is_empty() {
yield ProviderStreamChunk {
content,
reasoning_content: None,
finish_reason: None,
tool_calls: None,
model: model.clone(),
usage: None,
};
}
}
}
Ok(_) => continue,
Err(e) => {
// Attempt to probe for the actual error body
let probe_resp = probe_client
.post(&url)
.header("Authorization", format!("Bearer {}", api_key))
.json(&probe_body)
.send()
.await;
match probe_resp {
Ok(r) if !r.status().is_success() => {
let status = r.status();
let error_body = r.text().await.unwrap_or_default();
tracing::error!("OpenAI Responses Stream Error Probe ({}): {}", status, error_body);
Err(AppError::ProviderError(format!("OpenAI Responses API error ({}): {}", status, error_body)))?;
}
Ok(_) => {
// If the probe returned 200, but the stream ended, it might be a silent failure or timeout.
tracing::warn!("Responses stream ended prematurely (probe returned 200)");
Err(AppError::ProviderError(format!("Responses stream error (probe returned 200): {}", e)))?;
}
Err(probe_err) => {
tracing::error!("OpenAI Responses Stream Error Probe failed: {}", probe_err);
Err(AppError::ProviderError(format!("Responses stream error (probe failed: {}): {}", probe_err, e)))?;
}
}
}
}
}
};
Ok(Box::pin(stream))
}
}