1pub mod batches;
2pub mod responses;
3
4use anyhow::{Context as _, Result, anyhow};
5use futures::{AsyncBufReadExt, AsyncReadExt, StreamExt, io::BufReader, stream::BoxStream};
6use http_client::{
7 AsyncBody, HttpClient, Method, Request as HttpRequest, StatusCode,
8 http::{HeaderMap, HeaderValue},
9};
10use serde::{Deserialize, Serialize};
11use serde_json::Value;
12pub use settings::OpenAiReasoningEffort as ReasoningEffort;
13use std::{convert::TryFrom, future::Future};
14use strum::EnumIter;
15use thiserror::Error;
16
17pub const OPEN_AI_API_URL: &str = "https://api.openai.com/v1";
18
19fn is_none_or_empty<T: AsRef<[U]>, U>(opt: &Option<T>) -> bool {
20 opt.as_ref().is_none_or(|v| v.as_ref().is_empty())
21}
22
23#[derive(Clone, Copy, Serialize, Deserialize, Debug, Eq, PartialEq)]
24#[serde(rename_all = "lowercase")]
25pub enum Role {
26 User,
27 Assistant,
28 System,
29 Tool,
30}
31
32impl TryFrom<String> for Role {
33 type Error = anyhow::Error;
34
35 fn try_from(value: String) -> Result<Self> {
36 match value.as_str() {
37 "user" => Ok(Self::User),
38 "assistant" => Ok(Self::Assistant),
39 "system" => Ok(Self::System),
40 "tool" => Ok(Self::Tool),
41 _ => anyhow::bail!("invalid role '{value}'"),
42 }
43 }
44}
45
46impl From<Role> for String {
47 fn from(val: Role) -> Self {
48 match val {
49 Role::User => "user".to_owned(),
50 Role::Assistant => "assistant".to_owned(),
51 Role::System => "system".to_owned(),
52 Role::Tool => "tool".to_owned(),
53 }
54 }
55}
56
57#[cfg_attr(feature = "schemars", derive(schemars::JsonSchema))]
58#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq, EnumIter)]
59pub enum Model {
60 #[serde(rename = "gpt-3.5-turbo")]
61 ThreePointFiveTurbo,
62 #[serde(rename = "gpt-4")]
63 Four,
64 #[serde(rename = "gpt-4-turbo")]
65 FourTurbo,
66 #[serde(rename = "gpt-4o-mini")]
67 FourOmniMini,
68 #[serde(rename = "gpt-4.1-nano")]
69 FourPointOneNano,
70 #[serde(rename = "o1")]
71 O1,
72 #[serde(rename = "o3-mini")]
73 O3Mini,
74 #[serde(rename = "o3")]
75 O3,
76 #[serde(rename = "gpt-5")]
77 Five,
78 #[serde(rename = "gpt-5-codex")]
79 FiveCodex,
80 #[serde(rename = "gpt-5-mini")]
81 #[default]
82 FiveMini,
83 #[serde(rename = "gpt-5-nano")]
84 FiveNano,
85 #[serde(rename = "gpt-5.1")]
86 FivePointOne,
87 #[serde(rename = "gpt-5.2")]
88 FivePointTwo,
89 #[serde(rename = "gpt-5.2-codex")]
90 FivePointTwoCodex,
91 #[serde(rename = "gpt-5.3-codex")]
92 FivePointThreeCodex,
93 #[serde(rename = "custom")]
94 Custom {
95 name: String,
96 /// The name displayed in the UI, such as in the assistant panel model dropdown menu.
97 display_name: Option<String>,
98 max_tokens: u64,
99 max_output_tokens: Option<u64>,
100 max_completion_tokens: Option<u64>,
101 reasoning_effort: Option<ReasoningEffort>,
102 #[serde(default = "default_supports_chat_completions")]
103 supports_chat_completions: bool,
104 },
105}
106
107const fn default_supports_chat_completions() -> bool {
108 true
109}
110
111impl Model {
112 pub fn default_fast() -> Self {
113 Self::FiveMini
114 }
115
116 pub fn from_id(id: &str) -> Result<Self> {
117 match id {
118 "gpt-3.5-turbo" => Ok(Self::ThreePointFiveTurbo),
119 "gpt-4" => Ok(Self::Four),
120 "gpt-4-turbo-preview" => Ok(Self::FourTurbo),
121 "gpt-4o-mini" => Ok(Self::FourOmniMini),
122 "gpt-4.1-nano" => Ok(Self::FourPointOneNano),
123 "o1" => Ok(Self::O1),
124 "o3-mini" => Ok(Self::O3Mini),
125 "o3" => Ok(Self::O3),
126 "gpt-5" => Ok(Self::Five),
127 "gpt-5-codex" => Ok(Self::FiveCodex),
128 "gpt-5-mini" => Ok(Self::FiveMini),
129 "gpt-5-nano" => Ok(Self::FiveNano),
130 "gpt-5.1" => Ok(Self::FivePointOne),
131 "gpt-5.2" => Ok(Self::FivePointTwo),
132 "gpt-5.2-codex" => Ok(Self::FivePointTwoCodex),
133 "gpt-5.3-codex" => Ok(Self::FivePointThreeCodex),
134 invalid_id => anyhow::bail!("invalid model id '{invalid_id}'"),
135 }
136 }
137
138 pub fn id(&self) -> &str {
139 match self {
140 Self::ThreePointFiveTurbo => "gpt-3.5-turbo",
141 Self::Four => "gpt-4",
142 Self::FourTurbo => "gpt-4-turbo",
143 Self::FourOmniMini => "gpt-4o-mini",
144 Self::FourPointOneNano => "gpt-4.1-nano",
145 Self::O1 => "o1",
146 Self::O3Mini => "o3-mini",
147 Self::O3 => "o3",
148 Self::Five => "gpt-5",
149 Self::FiveCodex => "gpt-5-codex",
150 Self::FiveMini => "gpt-5-mini",
151 Self::FiveNano => "gpt-5-nano",
152 Self::FivePointOne => "gpt-5.1",
153 Self::FivePointTwo => "gpt-5.2",
154 Self::FivePointTwoCodex => "gpt-5.2-codex",
155 Self::FivePointThreeCodex => "gpt-5.3-codex",
156 Self::Custom { name, .. } => name,
157 }
158 }
159
160 pub fn display_name(&self) -> &str {
161 match self {
162 Self::ThreePointFiveTurbo => "gpt-3.5-turbo",
163 Self::Four => "gpt-4",
164 Self::FourTurbo => "gpt-4-turbo",
165 Self::FourOmniMini => "gpt-4o-mini",
166 Self::FourPointOneNano => "gpt-4.1-nano",
167 Self::O1 => "o1",
168 Self::O3Mini => "o3-mini",
169 Self::O3 => "o3",
170 Self::Five => "gpt-5",
171 Self::FiveCodex => "gpt-5-codex",
172 Self::FiveMini => "gpt-5-mini",
173 Self::FiveNano => "gpt-5-nano",
174 Self::FivePointOne => "gpt-5.1",
175 Self::FivePointTwo => "gpt-5.2",
176 Self::FivePointTwoCodex => "gpt-5.2-codex",
177 Self::FivePointThreeCodex => "gpt-5.3-codex",
178 Self::Custom { display_name, .. } => display_name.as_deref().unwrap_or(&self.id()),
179 }
180 }
181
182 pub fn max_token_count(&self) -> u64 {
183 match self {
184 Self::ThreePointFiveTurbo => 16_385,
185 Self::Four => 8_192,
186 Self::FourTurbo => 128_000,
187 Self::FourOmniMini => 128_000,
188 Self::FourPointOneNano => 1_047_576,
189 Self::O1 => 200_000,
190 Self::O3Mini => 200_000,
191 Self::O3 => 200_000,
192 Self::Five => 272_000,
193 Self::FiveCodex => 272_000,
194 Self::FiveMini => 272_000,
195 Self::FiveNano => 272_000,
196 Self::FivePointOne => 400_000,
197 Self::FivePointTwo => 400_000,
198 Self::FivePointTwoCodex => 400_000,
199 Self::FivePointThreeCodex => 400_000,
200 Self::Custom { max_tokens, .. } => *max_tokens,
201 }
202 }
203
204 pub fn max_output_tokens(&self) -> Option<u64> {
205 match self {
206 Self::Custom {
207 max_output_tokens, ..
208 } => *max_output_tokens,
209 Self::ThreePointFiveTurbo => Some(4_096),
210 Self::Four => Some(8_192),
211 Self::FourTurbo => Some(4_096),
212 Self::FourOmniMini => Some(16_384),
213 Self::FourPointOneNano => Some(32_768),
214 Self::O1 => Some(100_000),
215 Self::O3Mini => Some(100_000),
216 Self::O3 => Some(100_000),
217 Self::Five => Some(128_000),
218 Self::FiveCodex => Some(128_000),
219 Self::FiveMini => Some(128_000),
220 Self::FiveNano => Some(128_000),
221 Self::FivePointOne => Some(128_000),
222 Self::FivePointTwo => Some(128_000),
223 Self::FivePointTwoCodex => Some(128_000),
224 Self::FivePointThreeCodex => Some(128_000),
225 }
226 }
227
228 pub fn reasoning_effort(&self) -> Option<ReasoningEffort> {
229 match self {
230 Self::Custom {
231 reasoning_effort, ..
232 } => reasoning_effort.to_owned(),
233 Self::FivePointThreeCodex => Some(ReasoningEffort::Medium),
234 _ => None,
235 }
236 }
237
238 pub fn supports_chat_completions(&self) -> bool {
239 match self {
240 Self::Custom {
241 supports_chat_completions,
242 ..
243 } => *supports_chat_completions,
244 Self::FiveCodex | Self::FivePointTwoCodex | Self::FivePointThreeCodex => false,
245 _ => true,
246 }
247 }
248
249 /// Returns whether the given model supports the `parallel_tool_calls` parameter.
250 ///
251 /// If the model does not support the parameter, do not pass it up, or the API will return an error.
252 pub fn supports_parallel_tool_calls(&self) -> bool {
253 match self {
254 Self::ThreePointFiveTurbo
255 | Self::Four
256 | Self::FourTurbo
257 | Self::FourOmniMini
258 | Self::FourPointOneNano
259 | Self::Five
260 | Self::FiveCodex
261 | Self::FiveMini
262 | Self::FivePointOne
263 | Self::FivePointTwo
264 | Self::FivePointTwoCodex
265 | Self::FivePointThreeCodex
266 | Self::FiveNano => true,
267 Self::O1 | Self::O3 | Self::O3Mini | Model::Custom { .. } => false,
268 }
269 }
270
271 /// Returns whether the given model supports the `prompt_cache_key` parameter.
272 ///
273 /// If the model does not support the parameter, do not pass it up.
274 pub fn supports_prompt_cache_key(&self) -> bool {
275 true
276 }
277}
278
279#[derive(Debug, Serialize, Deserialize)]
280pub struct Request {
281 pub model: String,
282 pub messages: Vec<RequestMessage>,
283 pub stream: bool,
284 #[serde(default, skip_serializing_if = "Option::is_none")]
285 pub max_completion_tokens: Option<u64>,
286 #[serde(default, skip_serializing_if = "Vec::is_empty")]
287 pub stop: Vec<String>,
288 #[serde(default, skip_serializing_if = "Option::is_none")]
289 pub temperature: Option<f32>,
290 #[serde(default, skip_serializing_if = "Option::is_none")]
291 pub tool_choice: Option<ToolChoice>,
292 /// Whether to enable parallel function calling during tool use.
293 #[serde(default, skip_serializing_if = "Option::is_none")]
294 pub parallel_tool_calls: Option<bool>,
295 #[serde(default, skip_serializing_if = "Vec::is_empty")]
296 pub tools: Vec<ToolDefinition>,
297 #[serde(default, skip_serializing_if = "Option::is_none")]
298 pub prompt_cache_key: Option<String>,
299 #[serde(default, skip_serializing_if = "Option::is_none")]
300 pub reasoning_effort: Option<ReasoningEffort>,
301}
302
303#[derive(Debug, Serialize, Deserialize)]
304#[serde(rename_all = "lowercase")]
305pub enum ToolChoice {
306 Auto,
307 Required,
308 None,
309 #[serde(untagged)]
310 Other(ToolDefinition),
311}
312
313#[derive(Clone, Deserialize, Serialize, Debug)]
314#[serde(tag = "type", rename_all = "snake_case")]
315pub enum ToolDefinition {
316 #[allow(dead_code)]
317 Function { function: FunctionDefinition },
318}
319
320#[derive(Clone, Debug, Serialize, Deserialize)]
321pub struct FunctionDefinition {
322 pub name: String,
323 pub description: Option<String>,
324 pub parameters: Option<Value>,
325}
326
327#[derive(Clone, Serialize, Deserialize, Debug, Eq, PartialEq)]
328#[serde(tag = "role", rename_all = "lowercase")]
329pub enum RequestMessage {
330 Assistant {
331 content: Option<MessageContent>,
332 #[serde(default, skip_serializing_if = "Vec::is_empty")]
333 tool_calls: Vec<ToolCall>,
334 },
335 User {
336 content: MessageContent,
337 },
338 System {
339 content: MessageContent,
340 },
341 Tool {
342 content: MessageContent,
343 tool_call_id: String,
344 },
345}
346
347#[derive(Serialize, Deserialize, Clone, Debug, Eq, PartialEq)]
348#[serde(untagged)]
349pub enum MessageContent {
350 Plain(String),
351 Multipart(Vec<MessagePart>),
352}
353
354impl MessageContent {
355 pub fn empty() -> Self {
356 MessageContent::Multipart(vec![])
357 }
358
359 pub fn push_part(&mut self, part: MessagePart) {
360 match self {
361 MessageContent::Plain(text) => {
362 *self =
363 MessageContent::Multipart(vec![MessagePart::Text { text: text.clone() }, part]);
364 }
365 MessageContent::Multipart(parts) if parts.is_empty() => match part {
366 MessagePart::Text { text } => *self = MessageContent::Plain(text),
367 MessagePart::Image { .. } => *self = MessageContent::Multipart(vec![part]),
368 },
369 MessageContent::Multipart(parts) => parts.push(part),
370 }
371 }
372}
373
374impl From<Vec<MessagePart>> for MessageContent {
375 fn from(mut parts: Vec<MessagePart>) -> Self {
376 if let [MessagePart::Text { text }] = parts.as_mut_slice() {
377 MessageContent::Plain(std::mem::take(text))
378 } else {
379 MessageContent::Multipart(parts)
380 }
381 }
382}
383
384#[derive(Serialize, Deserialize, Clone, Debug, Eq, PartialEq)]
385#[serde(tag = "type")]
386pub enum MessagePart {
387 #[serde(rename = "text")]
388 Text { text: String },
389 #[serde(rename = "image_url")]
390 Image { image_url: ImageUrl },
391}
392
393#[derive(Serialize, Deserialize, Clone, Debug, Eq, PartialEq)]
394pub struct ImageUrl {
395 pub url: String,
396 #[serde(skip_serializing_if = "Option::is_none")]
397 pub detail: Option<String>,
398}
399
400#[derive(Clone, Serialize, Deserialize, Debug, Eq, PartialEq)]
401pub struct ToolCall {
402 pub id: String,
403 #[serde(flatten)]
404 pub content: ToolCallContent,
405}
406
407#[derive(Clone, Serialize, Deserialize, Debug, Eq, PartialEq)]
408#[serde(tag = "type", rename_all = "lowercase")]
409pub enum ToolCallContent {
410 Function { function: FunctionContent },
411}
412
413#[derive(Clone, Serialize, Deserialize, Debug, Eq, PartialEq)]
414pub struct FunctionContent {
415 pub name: String,
416 pub arguments: String,
417}
418
419#[derive(Clone, Serialize, Deserialize, Debug)]
420pub struct Response {
421 pub id: String,
422 pub object: String,
423 pub created: u64,
424 pub model: String,
425 pub choices: Vec<Choice>,
426 pub usage: Usage,
427}
428
429#[derive(Clone, Serialize, Deserialize, Debug)]
430pub struct Choice {
431 pub index: u32,
432 pub message: RequestMessage,
433 pub finish_reason: Option<String>,
434}
435
436#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
437pub struct ResponseMessageDelta {
438 pub role: Option<Role>,
439 pub content: Option<String>,
440 #[serde(default, skip_serializing_if = "is_none_or_empty")]
441 pub tool_calls: Option<Vec<ToolCallChunk>>,
442 #[serde(default, skip_serializing_if = "is_none_or_empty")]
443 pub reasoning_content: Option<String>,
444}
445
446#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
447pub struct ToolCallChunk {
448 pub index: usize,
449 pub id: Option<String>,
450
451 // There is also an optional `type` field that would determine if a
452 // function is there. Sometimes this streams in with the `function` before
453 // it streams in the `type`
454 pub function: Option<FunctionChunk>,
455}
456
457#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
458pub struct FunctionChunk {
459 pub name: Option<String>,
460 pub arguments: Option<String>,
461}
462
463#[derive(Clone, Serialize, Deserialize, Debug)]
464pub struct Usage {
465 pub prompt_tokens: u64,
466 pub completion_tokens: u64,
467 pub total_tokens: u64,
468}
469
470#[derive(Serialize, Deserialize, Debug)]
471pub struct ChoiceDelta {
472 pub index: u32,
473 pub delta: Option<ResponseMessageDelta>,
474 pub finish_reason: Option<String>,
475}
476
477#[derive(Error, Debug)]
478pub enum RequestError {
479 #[error("HTTP response error from {provider}'s API: status {status_code} - {body:?}")]
480 HttpResponseError {
481 provider: String,
482 status_code: StatusCode,
483 body: String,
484 headers: HeaderMap<HeaderValue>,
485 },
486 #[error(transparent)]
487 Other(#[from] anyhow::Error),
488}
489
490#[derive(Serialize, Deserialize, Debug)]
491pub struct ResponseStreamError {
492 message: String,
493}
494
495#[derive(Serialize, Deserialize, Debug)]
496#[serde(untagged)]
497pub enum ResponseStreamResult {
498 Ok(ResponseStreamEvent),
499 Err { error: ResponseStreamError },
500}
501
502#[derive(Serialize, Deserialize, Debug)]
503pub struct ResponseStreamEvent {
504 pub choices: Vec<ChoiceDelta>,
505 pub usage: Option<Usage>,
506}
507
508pub async fn non_streaming_completion(
509 client: &dyn HttpClient,
510 api_url: &str,
511 api_key: &str,
512 request: Request,
513) -> Result<Response, RequestError> {
514 let uri = format!("{api_url}/chat/completions");
515 let request_builder = HttpRequest::builder()
516 .method(Method::POST)
517 .uri(uri)
518 .header("Content-Type", "application/json")
519 .header("Authorization", format!("Bearer {}", api_key.trim()));
520
521 let request = request_builder
522 .body(AsyncBody::from(
523 serde_json::to_string(&request).map_err(|e| RequestError::Other(e.into()))?,
524 ))
525 .map_err(|e| RequestError::Other(e.into()))?;
526
527 let mut response = client.send(request).await?;
528 if response.status().is_success() {
529 let mut body = String::new();
530 response
531 .body_mut()
532 .read_to_string(&mut body)
533 .await
534 .map_err(|e| RequestError::Other(e.into()))?;
535
536 serde_json::from_str(&body).map_err(|e| RequestError::Other(e.into()))
537 } else {
538 let mut body = String::new();
539 response
540 .body_mut()
541 .read_to_string(&mut body)
542 .await
543 .map_err(|e| RequestError::Other(e.into()))?;
544
545 Err(RequestError::HttpResponseError {
546 provider: "openai".to_owned(),
547 status_code: response.status(),
548 body,
549 headers: response.headers().clone(),
550 })
551 }
552}
553
554pub async fn stream_completion(
555 client: &dyn HttpClient,
556 provider_name: &str,
557 api_url: &str,
558 api_key: &str,
559 request: Request,
560) -> Result<BoxStream<'static, Result<ResponseStreamEvent>>, RequestError> {
561 let uri = format!("{api_url}/chat/completions");
562 let request_builder = HttpRequest::builder()
563 .method(Method::POST)
564 .uri(uri)
565 .header("Content-Type", "application/json")
566 .header("Authorization", format!("Bearer {}", api_key.trim()));
567
568 let request = request_builder
569 .body(AsyncBody::from(
570 serde_json::to_string(&request).map_err(|e| RequestError::Other(e.into()))?,
571 ))
572 .map_err(|e| RequestError::Other(e.into()))?;
573
574 let mut response = client.send(request).await?;
575 if response.status().is_success() {
576 let reader = BufReader::new(response.into_body());
577 Ok(reader
578 .lines()
579 .filter_map(|line| async move {
580 match line {
581 Ok(line) => {
582 let line = line.strip_prefix("data: ").or_else(|| line.strip_prefix("data:"))?;
583 if line == "[DONE]" {
584 None
585 } else {
586 match serde_json::from_str(line) {
587 Ok(ResponseStreamResult::Ok(response)) => Some(Ok(response)),
588 Ok(ResponseStreamResult::Err { error }) => {
589 Some(Err(anyhow!(error.message)))
590 }
591 Err(error) => {
592 log::error!(
593 "Failed to parse OpenAI response into ResponseStreamResult: `{}`\n\
594 Response: `{}`",
595 error,
596 line,
597 );
598 Some(Err(anyhow!(error)))
599 }
600 }
601 }
602 }
603 Err(error) => Some(Err(anyhow!(error))),
604 }
605 })
606 .boxed())
607 } else {
608 let mut body = String::new();
609 response
610 .body_mut()
611 .read_to_string(&mut body)
612 .await
613 .map_err(|e| RequestError::Other(e.into()))?;
614
615 Err(RequestError::HttpResponseError {
616 provider: provider_name.to_owned(),
617 status_code: response.status(),
618 body,
619 headers: response.headers().clone(),
620 })
621 }
622}
623
624#[derive(Copy, Clone, Serialize, Deserialize)]
625pub enum OpenAiEmbeddingModel {
626 #[serde(rename = "text-embedding-3-small")]
627 TextEmbedding3Small,
628 #[serde(rename = "text-embedding-3-large")]
629 TextEmbedding3Large,
630}
631
632#[derive(Serialize)]
633struct OpenAiEmbeddingRequest<'a> {
634 model: OpenAiEmbeddingModel,
635 input: Vec<&'a str>,
636}
637
638#[derive(Deserialize)]
639pub struct OpenAiEmbeddingResponse {
640 pub data: Vec<OpenAiEmbedding>,
641}
642
643#[derive(Deserialize)]
644pub struct OpenAiEmbedding {
645 pub embedding: Vec<f32>,
646}
647
648pub fn embed<'a>(
649 client: &dyn HttpClient,
650 api_url: &str,
651 api_key: &str,
652 model: OpenAiEmbeddingModel,
653 texts: impl IntoIterator<Item = &'a str>,
654) -> impl 'static + Future<Output = Result<OpenAiEmbeddingResponse>> {
655 let uri = format!("{api_url}/embeddings");
656
657 let request = OpenAiEmbeddingRequest {
658 model,
659 input: texts.into_iter().collect(),
660 };
661 let body = AsyncBody::from(serde_json::to_string(&request).unwrap());
662 let request = HttpRequest::builder()
663 .method(Method::POST)
664 .uri(uri)
665 .header("Content-Type", "application/json")
666 .header("Authorization", format!("Bearer {}", api_key.trim()))
667 .body(body)
668 .map(|request| client.send(request));
669
670 async move {
671 let mut response = request?.await?;
672 let mut body = String::new();
673 response.body_mut().read_to_string(&mut body).await?;
674
675 anyhow::ensure!(
676 response.status().is_success(),
677 "error during embedding, status: {:?}, body: {:?}",
678 response.status(),
679 body
680 );
681 let response: OpenAiEmbeddingResponse =
682 serde_json::from_str(&body).context("failed to parse OpenAI embedding response")?;
683 Ok(response)
684 }
685}