open_ai.rs

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