open_ai.rs

  1use anyhow::{Context as _, Result, anyhow};
  2use futures::{AsyncBufReadExt, AsyncReadExt, StreamExt, io::BufReader, stream::BoxStream};
  3use http_client::{
  4    AsyncBody, HttpClient, Method, Request as HttpRequest, StatusCode,
  5    http::{HeaderMap, HeaderValue},
  6};
  7use serde::{Deserialize, Serialize};
  8use serde_json::Value;
  9pub use settings::OpenAiReasoningEffort as ReasoningEffort;
 10use std::{convert::TryFrom, future::Future};
 11use strum::EnumIter;
 12use thiserror::Error;
 13
 14pub const OPEN_AI_API_URL: &str = "https://api.openai.com/v1";
 15
 16fn is_none_or_empty<T: AsRef<[U]>, U>(opt: &Option<T>) -> bool {
 17    opt.as_ref().is_none_or(|v| v.as_ref().is_empty())
 18}
 19
 20#[derive(Clone, Copy, Serialize, Deserialize, Debug, Eq, PartialEq)]
 21#[serde(rename_all = "lowercase")]
 22pub enum Role {
 23    User,
 24    Assistant,
 25    System,
 26    Tool,
 27}
 28
 29impl TryFrom<String> for Role {
 30    type Error = anyhow::Error;
 31
 32    fn try_from(value: String) -> Result<Self> {
 33        match value.as_str() {
 34            "user" => Ok(Self::User),
 35            "assistant" => Ok(Self::Assistant),
 36            "system" => Ok(Self::System),
 37            "tool" => Ok(Self::Tool),
 38            _ => anyhow::bail!("invalid role '{value}'"),
 39        }
 40    }
 41}
 42
 43impl From<Role> for String {
 44    fn from(val: Role) -> Self {
 45        match val {
 46            Role::User => "user".to_owned(),
 47            Role::Assistant => "assistant".to_owned(),
 48            Role::System => "system".to_owned(),
 49            Role::Tool => "tool".to_owned(),
 50        }
 51    }
 52}
 53
 54#[cfg_attr(feature = "schemars", derive(schemars::JsonSchema))]
 55#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq, EnumIter)]
 56pub enum Model {
 57    #[serde(rename = "gpt-3.5-turbo")]
 58    ThreePointFiveTurbo,
 59    #[serde(rename = "gpt-4")]
 60    Four,
 61    #[serde(rename = "gpt-4-turbo")]
 62    FourTurbo,
 63    #[serde(rename = "gpt-4o")]
 64    #[default]
 65    FourOmni,
 66    #[serde(rename = "gpt-4o-mini")]
 67    FourOmniMini,
 68    #[serde(rename = "gpt-4.1")]
 69    FourPointOne,
 70    #[serde(rename = "gpt-4.1-mini")]
 71    FourPointOneMini,
 72    #[serde(rename = "gpt-4.1-nano")]
 73    FourPointOneNano,
 74    #[serde(rename = "o1")]
 75    O1,
 76    #[serde(rename = "o3-mini")]
 77    O3Mini,
 78    #[serde(rename = "o3")]
 79    O3,
 80    #[serde(rename = "o4-mini")]
 81    O4Mini,
 82    #[serde(rename = "gpt-5")]
 83    Five,
 84    #[serde(rename = "gpt-5-mini")]
 85    FiveMini,
 86    #[serde(rename = "gpt-5-nano")]
 87    FiveNano,
 88
 89    #[serde(rename = "custom")]
 90    Custom {
 91        name: String,
 92        /// The name displayed in the UI, such as in the assistant panel model dropdown menu.
 93        display_name: Option<String>,
 94        max_tokens: u64,
 95        max_output_tokens: Option<u64>,
 96        max_completion_tokens: Option<u64>,
 97        reasoning_effort: Option<ReasoningEffort>,
 98    },
 99}
100
101impl Model {
102    pub fn default_fast() -> Self {
103        // TODO: Replace with FiveMini since all other models are deprecated
104        Self::FourPointOneMini
105    }
106
107    pub fn from_id(id: &str) -> Result<Self> {
108        match id {
109            "gpt-3.5-turbo" => Ok(Self::ThreePointFiveTurbo),
110            "gpt-4" => Ok(Self::Four),
111            "gpt-4-turbo-preview" => Ok(Self::FourTurbo),
112            "gpt-4o" => Ok(Self::FourOmni),
113            "gpt-4o-mini" => Ok(Self::FourOmniMini),
114            "gpt-4.1" => Ok(Self::FourPointOne),
115            "gpt-4.1-mini" => Ok(Self::FourPointOneMini),
116            "gpt-4.1-nano" => Ok(Self::FourPointOneNano),
117            "o1" => Ok(Self::O1),
118            "o3-mini" => Ok(Self::O3Mini),
119            "o3" => Ok(Self::O3),
120            "o4-mini" => Ok(Self::O4Mini),
121            "gpt-5" => Ok(Self::Five),
122            "gpt-5-mini" => Ok(Self::FiveMini),
123            "gpt-5-nano" => Ok(Self::FiveNano),
124            invalid_id => anyhow::bail!("invalid model id '{invalid_id}'"),
125        }
126    }
127
128    pub fn id(&self) -> &str {
129        match self {
130            Self::ThreePointFiveTurbo => "gpt-3.5-turbo",
131            Self::Four => "gpt-4",
132            Self::FourTurbo => "gpt-4-turbo",
133            Self::FourOmni => "gpt-4o",
134            Self::FourOmniMini => "gpt-4o-mini",
135            Self::FourPointOne => "gpt-4.1",
136            Self::FourPointOneMini => "gpt-4.1-mini",
137            Self::FourPointOneNano => "gpt-4.1-nano",
138            Self::O1 => "o1",
139            Self::O3Mini => "o3-mini",
140            Self::O3 => "o3",
141            Self::O4Mini => "o4-mini",
142            Self::Five => "gpt-5",
143            Self::FiveMini => "gpt-5-mini",
144            Self::FiveNano => "gpt-5-nano",
145            Self::Custom { name, .. } => name,
146        }
147    }
148
149    pub fn display_name(&self) -> &str {
150        match self {
151            Self::ThreePointFiveTurbo => "gpt-3.5-turbo",
152            Self::Four => "gpt-4",
153            Self::FourTurbo => "gpt-4-turbo",
154            Self::FourOmni => "gpt-4o",
155            Self::FourOmniMini => "gpt-4o-mini",
156            Self::FourPointOne => "gpt-4.1",
157            Self::FourPointOneMini => "gpt-4.1-mini",
158            Self::FourPointOneNano => "gpt-4.1-nano",
159            Self::O1 => "o1",
160            Self::O3Mini => "o3-mini",
161            Self::O3 => "o3",
162            Self::O4Mini => "o4-mini",
163            Self::Five => "gpt-5",
164            Self::FiveMini => "gpt-5-mini",
165            Self::FiveNano => "gpt-5-nano",
166            Self::Custom {
167                name, display_name, ..
168            } => display_name.as_ref().unwrap_or(name),
169        }
170    }
171
172    pub fn max_token_count(&self) -> u64 {
173        match self {
174            Self::ThreePointFiveTurbo => 16_385,
175            Self::Four => 8_192,
176            Self::FourTurbo => 128_000,
177            Self::FourOmni => 128_000,
178            Self::FourOmniMini => 128_000,
179            Self::FourPointOne => 1_047_576,
180            Self::FourPointOneMini => 1_047_576,
181            Self::FourPointOneNano => 1_047_576,
182            Self::O1 => 200_000,
183            Self::O3Mini => 200_000,
184            Self::O3 => 200_000,
185            Self::O4Mini => 200_000,
186            Self::Five => 272_000,
187            Self::FiveMini => 272_000,
188            Self::FiveNano => 272_000,
189            Self::Custom { max_tokens, .. } => *max_tokens,
190        }
191    }
192
193    pub fn max_output_tokens(&self) -> Option<u64> {
194        match self {
195            Self::Custom {
196                max_output_tokens, ..
197            } => *max_output_tokens,
198            Self::ThreePointFiveTurbo => Some(4_096),
199            Self::Four => Some(8_192),
200            Self::FourTurbo => Some(4_096),
201            Self::FourOmni => Some(16_384),
202            Self::FourOmniMini => Some(16_384),
203            Self::FourPointOne => Some(32_768),
204            Self::FourPointOneMini => Some(32_768),
205            Self::FourPointOneNano => Some(32_768),
206            Self::O1 => Some(100_000),
207            Self::O3Mini => Some(100_000),
208            Self::O3 => Some(100_000),
209            Self::O4Mini => Some(100_000),
210            Self::Five => Some(128_000),
211            Self::FiveMini => Some(128_000),
212            Self::FiveNano => Some(128_000),
213        }
214    }
215
216    pub fn reasoning_effort(&self) -> Option<ReasoningEffort> {
217        match self {
218            Self::Custom {
219                reasoning_effort, ..
220            } => reasoning_effort.to_owned(),
221            _ => None,
222        }
223    }
224
225    /// Returns whether the given model supports the `parallel_tool_calls` parameter.
226    ///
227    /// If the model does not support the parameter, do not pass it up, or the API will return an error.
228    pub fn supports_parallel_tool_calls(&self) -> bool {
229        match self {
230            Self::ThreePointFiveTurbo
231            | Self::Four
232            | Self::FourTurbo
233            | Self::FourOmni
234            | Self::FourOmniMini
235            | Self::FourPointOne
236            | Self::FourPointOneMini
237            | Self::FourPointOneNano
238            | Self::Five
239            | Self::FiveMini
240            | Self::FiveNano => true,
241            Self::O1 | Self::O3 | Self::O3Mini | Self::O4Mini | Model::Custom { .. } => false,
242        }
243    }
244
245    /// Returns whether the given model supports the `prompt_cache_key` parameter.
246    ///
247    /// If the model does not support the parameter, do not pass it up.
248    pub fn supports_prompt_cache_key(&self) -> bool {
249        true
250    }
251}
252
253#[derive(Debug, Serialize, Deserialize)]
254pub struct Request {
255    pub model: String,
256    pub messages: Vec<RequestMessage>,
257    pub stream: bool,
258    #[serde(default, skip_serializing_if = "Option::is_none")]
259    pub max_completion_tokens: Option<u64>,
260    #[serde(default, skip_serializing_if = "Vec::is_empty")]
261    pub stop: Vec<String>,
262    pub temperature: f32,
263    #[serde(default, skip_serializing_if = "Option::is_none")]
264    pub tool_choice: Option<ToolChoice>,
265    /// Whether to enable parallel function calling during tool use.
266    #[serde(default, skip_serializing_if = "Option::is_none")]
267    pub parallel_tool_calls: Option<bool>,
268    #[serde(default, skip_serializing_if = "Vec::is_empty")]
269    pub tools: Vec<ToolDefinition>,
270    #[serde(default, skip_serializing_if = "Option::is_none")]
271    pub prompt_cache_key: Option<String>,
272    #[serde(default, skip_serializing_if = "Option::is_none")]
273    pub reasoning_effort: Option<ReasoningEffort>,
274}
275
276#[derive(Debug, Serialize, Deserialize)]
277#[serde(rename_all = "lowercase")]
278pub enum ToolChoice {
279    Auto,
280    Required,
281    None,
282    #[serde(untagged)]
283    Other(ToolDefinition),
284}
285
286#[derive(Clone, Deserialize, Serialize, Debug)]
287#[serde(tag = "type", rename_all = "snake_case")]
288pub enum ToolDefinition {
289    #[allow(dead_code)]
290    Function { function: FunctionDefinition },
291}
292
293#[derive(Clone, Debug, Serialize, Deserialize)]
294pub struct FunctionDefinition {
295    pub name: String,
296    pub description: Option<String>,
297    pub parameters: Option<Value>,
298}
299
300#[derive(Clone, Serialize, Deserialize, Debug, Eq, PartialEq)]
301#[serde(tag = "role", rename_all = "lowercase")]
302pub enum RequestMessage {
303    Assistant {
304        content: Option<MessageContent>,
305        #[serde(default, skip_serializing_if = "Vec::is_empty")]
306        tool_calls: Vec<ToolCall>,
307    },
308    User {
309        content: MessageContent,
310    },
311    System {
312        content: MessageContent,
313    },
314    Tool {
315        content: MessageContent,
316        tool_call_id: String,
317    },
318}
319
320#[derive(Serialize, Deserialize, Clone, Debug, Eq, PartialEq)]
321#[serde(untagged)]
322pub enum MessageContent {
323    Plain(String),
324    Multipart(Vec<MessagePart>),
325}
326
327impl MessageContent {
328    pub fn empty() -> Self {
329        MessageContent::Multipart(vec![])
330    }
331
332    pub fn push_part(&mut self, part: MessagePart) {
333        match self {
334            MessageContent::Plain(text) => {
335                *self =
336                    MessageContent::Multipart(vec![MessagePart::Text { text: text.clone() }, part]);
337            }
338            MessageContent::Multipart(parts) if parts.is_empty() => match part {
339                MessagePart::Text { text } => *self = MessageContent::Plain(text),
340                MessagePart::Image { .. } => *self = MessageContent::Multipart(vec![part]),
341            },
342            MessageContent::Multipart(parts) => parts.push(part),
343        }
344    }
345}
346
347impl From<Vec<MessagePart>> for MessageContent {
348    fn from(mut parts: Vec<MessagePart>) -> Self {
349        if let [MessagePart::Text { text }] = parts.as_mut_slice() {
350            MessageContent::Plain(std::mem::take(text))
351        } else {
352            MessageContent::Multipart(parts)
353        }
354    }
355}
356
357#[derive(Serialize, Deserialize, Clone, Debug, Eq, PartialEq)]
358#[serde(tag = "type")]
359pub enum MessagePart {
360    #[serde(rename = "text")]
361    Text { text: String },
362    #[serde(rename = "image_url")]
363    Image { image_url: ImageUrl },
364}
365
366#[derive(Serialize, Deserialize, Clone, Debug, Eq, PartialEq)]
367pub struct ImageUrl {
368    pub url: String,
369    #[serde(skip_serializing_if = "Option::is_none")]
370    pub detail: Option<String>,
371}
372
373#[derive(Clone, Serialize, Deserialize, Debug, Eq, PartialEq)]
374pub struct ToolCall {
375    pub id: String,
376    #[serde(flatten)]
377    pub content: ToolCallContent,
378}
379
380#[derive(Clone, Serialize, Deserialize, Debug, Eq, PartialEq)]
381#[serde(tag = "type", rename_all = "lowercase")]
382pub enum ToolCallContent {
383    Function { function: FunctionContent },
384}
385
386#[derive(Clone, Serialize, Deserialize, Debug, Eq, PartialEq)]
387pub struct FunctionContent {
388    pub name: String,
389    pub arguments: String,
390}
391
392#[derive(Clone, Serialize, Deserialize, Debug)]
393pub struct Response {
394    pub id: String,
395    pub object: String,
396    pub created: u64,
397    pub model: String,
398    pub choices: Vec<Choice>,
399    pub usage: Usage,
400}
401
402#[derive(Clone, Serialize, Deserialize, Debug)]
403pub struct Choice {
404    pub index: u32,
405    pub message: RequestMessage,
406    pub finish_reason: Option<String>,
407}
408
409#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
410pub struct ResponseMessageDelta {
411    pub role: Option<Role>,
412    pub content: Option<String>,
413    #[serde(default, skip_serializing_if = "is_none_or_empty")]
414    pub tool_calls: Option<Vec<ToolCallChunk>>,
415}
416
417#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
418pub struct ToolCallChunk {
419    pub index: usize,
420    pub id: Option<String>,
421
422    // There is also an optional `type` field that would determine if a
423    // function is there. Sometimes this streams in with the `function` before
424    // it streams in the `type`
425    pub function: Option<FunctionChunk>,
426}
427
428#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
429pub struct FunctionChunk {
430    pub name: Option<String>,
431    pub arguments: Option<String>,
432}
433
434#[derive(Clone, Serialize, Deserialize, Debug)]
435pub struct Usage {
436    pub prompt_tokens: u64,
437    pub completion_tokens: u64,
438    pub total_tokens: u64,
439}
440
441#[derive(Serialize, Deserialize, Debug)]
442pub struct ChoiceDelta {
443    pub index: u32,
444    pub delta: Option<ResponseMessageDelta>,
445    pub finish_reason: Option<String>,
446}
447
448#[derive(Error, Debug)]
449pub enum RequestError {
450    #[error("HTTP response error from {provider}'s API: status {status_code} - {body:?}")]
451    HttpResponseError {
452        provider: String,
453        status_code: StatusCode,
454        body: String,
455        headers: HeaderMap<HeaderValue>,
456    },
457    #[error(transparent)]
458    Other(#[from] anyhow::Error),
459}
460
461#[derive(Serialize, Deserialize, Debug)]
462pub struct ResponseStreamError {
463    message: String,
464}
465
466#[derive(Serialize, Deserialize, Debug)]
467#[serde(untagged)]
468pub enum ResponseStreamResult {
469    Ok(ResponseStreamEvent),
470    Err { error: ResponseStreamError },
471}
472
473#[derive(Serialize, Deserialize, Debug)]
474pub struct ResponseStreamEvent {
475    pub choices: Vec<ChoiceDelta>,
476    pub usage: Option<Usage>,
477}
478
479pub async fn stream_completion(
480    client: &dyn HttpClient,
481    provider_name: &str,
482    api_url: &str,
483    api_key: &str,
484    request: Request,
485) -> Result<BoxStream<'static, Result<ResponseStreamEvent>>, RequestError> {
486    let uri = format!("{api_url}/chat/completions");
487    let request_builder = HttpRequest::builder()
488        .method(Method::POST)
489        .uri(uri)
490        .header("Content-Type", "application/json")
491        .header("Authorization", format!("Bearer {}", api_key.trim()));
492
493    let request = request_builder
494        .body(AsyncBody::from(
495            serde_json::to_string(&request).map_err(|e| RequestError::Other(e.into()))?,
496        ))
497        .map_err(|e| RequestError::Other(e.into()))?;
498
499    let mut response = client.send(request).await?;
500    if response.status().is_success() {
501        let reader = BufReader::new(response.into_body());
502        Ok(reader
503            .lines()
504            .filter_map(|line| async move {
505                match line {
506                    Ok(line) => {
507                        let line = line.strip_prefix("data: ").or_else(|| line.strip_prefix("data:"))?;
508                        if line == "[DONE]" {
509                            None
510                        } else {
511                            match serde_json::from_str(line) {
512                                Ok(ResponseStreamResult::Ok(response)) => Some(Ok(response)),
513                                Ok(ResponseStreamResult::Err { error }) => {
514                                    Some(Err(anyhow!(error.message)))
515                                }
516                                Err(error) => {
517                                    log::error!(
518                                        "Failed to parse OpenAI response into ResponseStreamResult: `{}`\n\
519                                        Response: `{}`",
520                                        error,
521                                        line,
522                                    );
523                                    Some(Err(anyhow!(error)))
524                                }
525                            }
526                        }
527                    }
528                    Err(error) => Some(Err(anyhow!(error))),
529                }
530            })
531            .boxed())
532    } else {
533        let mut body = String::new();
534        response
535            .body_mut()
536            .read_to_string(&mut body)
537            .await
538            .map_err(|e| RequestError::Other(e.into()))?;
539
540        Err(RequestError::HttpResponseError {
541            provider: provider_name.to_owned(),
542            status_code: response.status(),
543            body,
544            headers: response.headers().clone(),
545        })
546    }
547}
548
549#[derive(Copy, Clone, Serialize, Deserialize)]
550pub enum OpenAiEmbeddingModel {
551    #[serde(rename = "text-embedding-3-small")]
552    TextEmbedding3Small,
553    #[serde(rename = "text-embedding-3-large")]
554    TextEmbedding3Large,
555}
556
557#[derive(Serialize)]
558struct OpenAiEmbeddingRequest<'a> {
559    model: OpenAiEmbeddingModel,
560    input: Vec<&'a str>,
561}
562
563#[derive(Deserialize)]
564pub struct OpenAiEmbeddingResponse {
565    pub data: Vec<OpenAiEmbedding>,
566}
567
568#[derive(Deserialize)]
569pub struct OpenAiEmbedding {
570    pub embedding: Vec<f32>,
571}
572
573pub fn embed<'a>(
574    client: &dyn HttpClient,
575    api_url: &str,
576    api_key: &str,
577    model: OpenAiEmbeddingModel,
578    texts: impl IntoIterator<Item = &'a str>,
579) -> impl 'static + Future<Output = Result<OpenAiEmbeddingResponse>> {
580    let uri = format!("{api_url}/embeddings");
581
582    let request = OpenAiEmbeddingRequest {
583        model,
584        input: texts.into_iter().collect(),
585    };
586    let body = AsyncBody::from(serde_json::to_string(&request).unwrap());
587    let request = HttpRequest::builder()
588        .method(Method::POST)
589        .uri(uri)
590        .header("Content-Type", "application/json")
591        .header("Authorization", format!("Bearer {}", api_key.trim()))
592        .body(body)
593        .map(|request| client.send(request));
594
595    async move {
596        let mut response = request?.await?;
597        let mut body = String::new();
598        response.body_mut().read_to_string(&mut body).await?;
599
600        anyhow::ensure!(
601            response.status().is_success(),
602            "error during embedding, status: {:?}, body: {:?}",
603            response.status(),
604            body
605        );
606        let response: OpenAiEmbeddingResponse =
607            serde_json::from_str(&body).context("failed to parse OpenAI embedding response")?;
608        Ok(response)
609    }
610}