1use std::fmt::Debug;
26
27use super::openai::{TranscriptionResponse, send_compatible_streaming_request};
28use crate::client::{
29 self, ApiKey, Capabilities, Capable, DebugExt, Nothing, Provider, ProviderBuilder,
30 ProviderClient,
31};
32use crate::completion::GetTokenUsage;
33use crate::http_client::multipart::Part;
34use crate::http_client::{self, HttpClientExt, MultipartForm, bearer_auth_header};
35use crate::streaming::StreamingCompletionResponse;
36use crate::transcription::TranscriptionError;
37use crate::{
38 completion::{self, CompletionError, CompletionRequest},
39 embeddings::{self, EmbeddingError},
40 json_utils,
41 providers::openai,
42 telemetry::SpanCombinator,
43 transcription::{self},
44};
45use bytes::Bytes;
46use serde::{Deserialize, Serialize};
47use serde_json::json;
48const DEFAULT_API_VERSION: &str = "2024-10-21";
53
54#[derive(Debug, Clone)]
55pub struct AzureExt {
56 endpoint: String,
57 api_version: String,
58}
59
60impl DebugExt for AzureExt {
61 fn fields(&self) -> impl Iterator<Item = (&'static str, &dyn std::fmt::Debug)> {
62 [
63 ("endpoint", (&self.endpoint as &dyn Debug)),
64 ("api_version", (&self.api_version as &dyn Debug)),
65 ]
66 .into_iter()
67 }
68}
69
70#[derive(Debug, Clone)]
75pub struct AzureExtBuilder {
76 endpoint: Option<String>,
77 api_version: String,
78}
79
80impl Default for AzureExtBuilder {
81 fn default() -> Self {
82 Self {
83 endpoint: None,
84 api_version: DEFAULT_API_VERSION.into(),
85 }
86 }
87}
88
89pub type Client<H = reqwest::Client> = client::Client<AzureExt, H>;
90pub type ClientBuilder<H = crate::markers::Missing> =
91 client::ClientBuilder<AzureExtBuilder, AzureOpenAIAuth, H>;
92
93impl Provider for AzureExt {
94 type Builder = AzureExtBuilder;
95
96 const VERIFY_PATH: &'static str = "";
98}
99
100impl<H> Capabilities<H> for AzureExt {
101 type Completion = Capable<CompletionModel<H>>;
102 type Embeddings = Capable<EmbeddingModel<H>>;
103 type Transcription = Capable<TranscriptionModel<H>>;
104 type ModelListing = Nothing;
105 #[cfg(feature = "image")]
106 type ImageGeneration = Nothing;
107 #[cfg(feature = "audio")]
108 type AudioGeneration = Capable<AudioGenerationModel<H>>;
109}
110
111impl ProviderBuilder for AzureExtBuilder {
112 type Extension<H>
113 = AzureExt
114 where
115 H: HttpClientExt;
116 type ApiKey = AzureOpenAIAuth;
117
118 const BASE_URL: &'static str = "";
119
120 fn build<H>(
121 builder: &client::ClientBuilder<Self, Self::ApiKey, H>,
122 ) -> http_client::Result<Self::Extension<H>>
123 where
124 H: HttpClientExt,
125 {
126 let AzureExtBuilder {
127 endpoint,
128 api_version,
129 ..
130 } = builder.ext().clone();
131
132 match endpoint {
133 Some(endpoint) => Ok(AzureExt {
134 endpoint,
135 api_version,
136 }),
137 None => Err(http_client::Error::Instance(
138 "Azure client must be provided an endpoint prior to building".into(),
139 )),
140 }
141 }
142
143 fn finish<H>(
144 &self,
145 mut builder: client::ClientBuilder<Self, Self::ApiKey, H>,
146 ) -> http_client::Result<client::ClientBuilder<Self, Self::ApiKey, H>> {
147 use AzureOpenAIAuth::*;
148
149 let auth = builder.get_api_key().clone();
150
151 match auth {
152 Token(token) => bearer_auth_header(builder.headers_mut(), token.as_str())?,
153 ApiKey(key) => {
154 let k = http::HeaderName::from_static("api-key");
155 let v = http::HeaderValue::from_str(key.as_str())?;
156
157 builder.headers_mut().insert(k, v);
158 }
159 }
160
161 Ok(builder)
162 }
163}
164
165impl<H> ClientBuilder<H> {
166 pub fn api_version(mut self, api_version: &str) -> Self {
168 self.ext_mut().api_version = api_version.into();
169
170 self
171 }
172}
173
174impl<H> client::ClientBuilder<AzureExtBuilder, AzureOpenAIAuth, H> {
175 pub fn azure_endpoint(self, endpoint: String) -> ClientBuilder<H> {
177 self.over_ext(|AzureExtBuilder { api_version, .. }| AzureExtBuilder {
178 endpoint: Some(endpoint),
179 api_version,
180 })
181 }
182}
183
184#[derive(Clone)]
187pub enum AzureOpenAIAuth {
188 ApiKey(String),
189 Token(String),
190}
191
192impl ApiKey for AzureOpenAIAuth {}
193
194impl std::fmt::Debug for AzureOpenAIAuth {
195 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
196 match self {
197 Self::ApiKey(_) => write!(f, "API key <REDACTED>"),
198 Self::Token(_) => write!(f, "Token <REDACTED>"),
199 }
200 }
201}
202
203impl<S> From<S> for AzureOpenAIAuth
204where
205 S: Into<String>,
206{
207 fn from(token: S) -> Self {
208 AzureOpenAIAuth::Token(token.into())
209 }
210}
211
212impl<T> Client<T>
213where
214 T: HttpClientExt,
215{
216 fn endpoint(&self) -> &str {
217 &self.ext().endpoint
218 }
219
220 fn api_version(&self) -> &str {
221 &self.ext().api_version
222 }
223
224 fn post_embedding(&self, deployment_id: &str) -> http_client::Result<http_client::Builder> {
225 let url = format!(
226 "{}/openai/deployments/{}/embeddings?api-version={}",
227 self.endpoint(),
228 deployment_id.trim_start_matches('/'),
229 self.api_version()
230 );
231
232 self.post(&url)
233 }
234
235 #[cfg(feature = "audio")]
236 fn post_audio_generation(
237 &self,
238 deployment_id: &str,
239 ) -> http_client::Result<http_client::Builder> {
240 let url = format!(
241 "{}/openai/deployments/{}/audio/speech?api-version={}",
242 self.endpoint(),
243 deployment_id.trim_start_matches('/'),
244 self.api_version()
245 );
246
247 self.post(url)
248 }
249
250 fn post_chat_completion(
251 &self,
252 deployment_id: &str,
253 ) -> http_client::Result<http_client::Builder> {
254 let url = format!(
255 "{}/openai/deployments/{}/chat/completions?api-version={}",
256 self.endpoint(),
257 deployment_id.trim_start_matches('/'),
258 self.api_version()
259 );
260
261 self.post(&url)
262 }
263
264 fn post_transcription(&self, deployment_id: &str) -> http_client::Result<http_client::Builder> {
265 let url = format!(
266 "{}/openai/deployments/{}/audio/translations?api-version={}",
267 self.endpoint(),
268 deployment_id.trim_start_matches('/'),
269 self.api_version()
270 );
271
272 self.post(&url)
273 }
274
275 #[cfg(feature = "image")]
276 fn post_image_generation(
277 &self,
278 deployment_id: &str,
279 ) -> http_client::Result<http_client::Builder> {
280 let url = format!(
281 "{}/openai/deployments/{}/images/generations?api-version={}",
282 self.endpoint(),
283 deployment_id.trim_start_matches('/'),
284 self.api_version()
285 );
286
287 self.post(&url)
288 }
289}
290
291pub struct AzureOpenAIClientParams {
292 api_key: String,
293 version: String,
294 header: String,
295}
296
297impl ProviderClient for Client {
298 type Input = AzureOpenAIClientParams;
299 type Error = crate::client::ProviderClientError;
300
301 fn from_env() -> Result<Self, Self::Error> {
303 let auth = if let Some(api_key) = crate::client::optional_env_var("AZURE_API_KEY")? {
304 AzureOpenAIAuth::ApiKey(api_key)
305 } else if let Some(token) = crate::client::optional_env_var("AZURE_TOKEN")? {
306 AzureOpenAIAuth::Token(token)
307 } else {
308 return Err(crate::client::ProviderClientError::InvalidConfiguration(
309 "either `AZURE_API_KEY` or `AZURE_TOKEN` must be set",
310 ));
311 };
312
313 let api_version = crate::client::required_env_var("AZURE_API_VERSION")?;
314 let azure_endpoint = crate::client::required_env_var("AZURE_ENDPOINT")?;
315
316 Self::builder()
317 .api_key(auth)
318 .azure_endpoint(azure_endpoint)
319 .api_version(&api_version)
320 .build()
321 .map_err(Into::into)
322 }
323
324 fn from_val(
325 AzureOpenAIClientParams {
326 api_key,
327 version,
328 header,
329 }: Self::Input,
330 ) -> Result<Self, Self::Error> {
331 let auth = AzureOpenAIAuth::ApiKey(api_key.to_string());
332
333 Self::builder()
334 .api_key(auth)
335 .azure_endpoint(header)
336 .api_version(&version)
337 .build()
338 .map_err(Into::into)
339 }
340}
341
342#[derive(Debug, Deserialize)]
343struct ApiErrorResponse {
344 message: String,
345}
346
347#[derive(Debug, Deserialize)]
348#[serde(untagged)]
349enum ApiResponse<T> {
350 Ok(T),
351 Err(ApiErrorResponse),
352}
353
354pub const TEXT_EMBEDDING_3_LARGE: &str = "text-embedding-3-large";
360pub const TEXT_EMBEDDING_3_SMALL: &str = "text-embedding-3-small";
362pub const TEXT_EMBEDDING_ADA_002: &str = "text-embedding-ada-002";
364
365fn model_dimensions_from_identifier(identifier: &str) -> Option<usize> {
366 match identifier {
367 TEXT_EMBEDDING_3_LARGE => Some(3_072),
368 TEXT_EMBEDDING_3_SMALL | TEXT_EMBEDDING_ADA_002 => Some(1_536),
369 _ => None,
370 }
371}
372
373#[derive(Debug, Deserialize)]
374pub struct EmbeddingResponse {
375 pub object: String,
376 pub data: Vec<EmbeddingData>,
377 pub model: String,
378 pub usage: Usage,
379}
380
381impl From<ApiErrorResponse> for EmbeddingError {
382 fn from(err: ApiErrorResponse) -> Self {
383 EmbeddingError::ProviderError(err.message)
384 }
385}
386
387impl From<ApiResponse<EmbeddingResponse>> for Result<EmbeddingResponse, EmbeddingError> {
388 fn from(value: ApiResponse<EmbeddingResponse>) -> Self {
389 match value {
390 ApiResponse::Ok(response) => Ok(response),
391 ApiResponse::Err(err) => Err(EmbeddingError::ProviderError(err.message)),
392 }
393 }
394}
395
396#[derive(Debug, Deserialize)]
397pub struct EmbeddingData {
398 pub object: String,
399 pub embedding: Vec<f64>,
400 pub index: usize,
401}
402
403#[derive(Clone, Debug, Deserialize)]
404pub struct Usage {
405 pub prompt_tokens: usize,
406 pub total_tokens: usize,
407}
408
409impl GetTokenUsage for Usage {
410 fn token_usage(&self) -> Option<crate::completion::Usage> {
411 let mut usage = crate::completion::Usage::new();
412
413 usage.input_tokens = self.prompt_tokens as u64;
414 usage.total_tokens = self.total_tokens as u64;
415 usage.output_tokens = usage.total_tokens - usage.input_tokens;
416
417 Some(usage)
418 }
419}
420
421impl std::fmt::Display for Usage {
422 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
423 write!(
424 f,
425 "Prompt tokens: {} Total tokens: {}",
426 self.prompt_tokens, self.total_tokens
427 )
428 }
429}
430
431#[derive(Clone)]
432pub struct EmbeddingModel<T = reqwest::Client> {
433 client: Client<T>,
434 pub model: String,
435 ndims: usize,
436}
437
438impl<T> embeddings::EmbeddingModel for EmbeddingModel<T>
439where
440 T: HttpClientExt + Default + Clone + 'static,
441{
442 const MAX_DOCUMENTS: usize = 1024;
443
444 type Client = Client<T>;
445
446 fn make(client: &Self::Client, model: impl Into<String>, dims: Option<usize>) -> Self {
447 Self::new(client.clone(), model, dims)
448 }
449
450 fn ndims(&self) -> usize {
451 self.ndims
452 }
453
454 async fn embed_texts(
455 &self,
456 documents: impl IntoIterator<Item = String>,
457 ) -> Result<Vec<embeddings::Embedding>, EmbeddingError> {
458 let documents = documents.into_iter().collect::<Vec<_>>();
459
460 let mut body = json!({
461 "input": documents,
462 });
463
464 let body_object = body.as_object_mut().ok_or_else(|| {
465 EmbeddingError::ResponseError("embedding request body must be a JSON object".into())
466 })?;
467
468 if self.ndims > 0 && self.model.as_str() != TEXT_EMBEDDING_ADA_002 {
469 body_object.insert("dimensions".to_owned(), json!(self.ndims));
470 }
471
472 let body = serde_json::to_vec(&body)?;
473
474 let req = self
475 .client
476 .post_embedding(self.model.as_str())?
477 .body(body)
478 .map_err(|e| EmbeddingError::HttpError(e.into()))?;
479
480 let response = self.client.send(req).await?;
481
482 if response.status().is_success() {
483 let body: Vec<u8> = response.into_body().await?;
484 let body: ApiResponse<EmbeddingResponse> = serde_json::from_slice(&body)?;
485
486 match body {
487 ApiResponse::Ok(response) => {
488 tracing::info!(target: "rig",
489 "Azure embedding token usage: {}",
490 response.usage
491 );
492
493 if response.data.len() != documents.len() {
494 return Err(EmbeddingError::ResponseError(
495 "Response data length does not match input length".into(),
496 ));
497 }
498
499 Ok(response
500 .data
501 .into_iter()
502 .zip(documents.into_iter())
503 .map(|(embedding, document)| embeddings::Embedding {
504 document,
505 vec: embedding.embedding,
506 })
507 .collect())
508 }
509 ApiResponse::Err(err) => Err(EmbeddingError::ProviderError(err.message)),
510 }
511 } else {
512 let text = http_client::text(response).await?;
513 Err(EmbeddingError::ProviderError(text))
514 }
515 }
516}
517
518impl<T> EmbeddingModel<T> {
519 pub fn new(client: Client<T>, model: impl Into<String>, ndims: Option<usize>) -> Self {
520 let model = model.into();
521 let ndims = ndims
522 .or(model_dimensions_from_identifier(&model))
523 .unwrap_or_default();
524
525 Self {
526 client,
527 model,
528 ndims,
529 }
530 }
531
532 pub fn with_model(client: Client<T>, model: &str, ndims: Option<usize>) -> Self {
533 let ndims = ndims.unwrap_or_default();
534
535 Self {
536 client,
537 model: model.into(),
538 ndims,
539 }
540 }
541}
542
543pub const O1: &str = "o1";
549pub const O1_PREVIEW: &str = "o1-preview";
551pub const O1_MINI: &str = "o1-mini";
553pub const GPT_4O: &str = "gpt-4o";
555pub const GPT_4O_MINI: &str = "gpt-4o-mini";
557pub const GPT_4O_REALTIME_PREVIEW: &str = "gpt-4o-realtime-preview";
559pub const GPT_4_TURBO: &str = "gpt-4";
561pub const GPT_4: &str = "gpt-4";
563pub const GPT_4_32K: &str = "gpt-4-32k";
565pub const GPT_4_32K_0613: &str = "gpt-4-32k";
567pub const GPT_35_TURBO: &str = "gpt-3.5-turbo";
569pub const GPT_35_TURBO_INSTRUCT: &str = "gpt-3.5-turbo-instruct";
571pub const GPT_35_TURBO_16K: &str = "gpt-3.5-turbo-16k";
573
574#[derive(Debug, Serialize, Deserialize)]
575pub(super) struct AzureOpenAICompletionRequest {
576 model: String,
577 pub messages: Vec<openai::Message>,
578 #[serde(skip_serializing_if = "Option::is_none")]
579 temperature: Option<f64>,
580 #[serde(skip_serializing_if = "Vec::is_empty")]
581 tools: Vec<openai::ToolDefinition>,
582 #[serde(skip_serializing_if = "Option::is_none")]
583 tool_choice: Option<crate::providers::openai::ToolChoice>,
584 #[serde(flatten, skip_serializing_if = "Option::is_none")]
585 pub additional_params: Option<serde_json::Value>,
586}
587
588impl TryFrom<(&str, CompletionRequest)> for AzureOpenAICompletionRequest {
589 type Error = CompletionError;
590
591 fn try_from((model, req): (&str, CompletionRequest)) -> Result<Self, Self::Error> {
592 let model = req.model.clone().unwrap_or_else(|| model.to_string());
593 if req.tool_choice.is_some() {
594 tracing::warn!("Tool choice is currently not supported in Azure OpenAI.");
595 }
596
597 let mut full_history: Vec<openai::Message> = match &req.preamble {
598 Some(preamble) => vec![openai::Message::system(preamble)],
599 None => vec![],
600 };
601
602 if let Some(docs) = req.normalized_documents() {
603 let docs: Vec<openai::Message> = docs.try_into()?;
604 full_history.extend(docs);
605 }
606
607 let chat_history: Vec<openai::Message> = req
608 .chat_history
609 .clone()
610 .into_iter()
611 .map(|message| message.try_into())
612 .collect::<Result<Vec<Vec<openai::Message>>, _>>()?
613 .into_iter()
614 .flatten()
615 .collect();
616
617 full_history.extend(chat_history);
618
619 let tool_choice = req
620 .tool_choice
621 .clone()
622 .map(crate::providers::openai::ToolChoice::try_from)
623 .transpose()?;
624
625 let additional_params = if let Some(schema) = req.output_schema {
626 let name = schema
627 .as_object()
628 .and_then(|o| o.get("title"))
629 .and_then(|v| v.as_str())
630 .unwrap_or("response_schema")
631 .to_string();
632 let mut schema_value = schema.to_value();
633 openai::sanitize_schema(&mut schema_value);
634 let response_format = serde_json::json!({
635 "response_format": {
636 "type": "json_schema",
637 "json_schema": {
638 "name": name,
639 "strict": true,
640 "schema": schema_value
641 }
642 }
643 });
644 Some(match req.additional_params {
645 Some(existing) => json_utils::merge(existing, response_format),
646 None => response_format,
647 })
648 } else {
649 req.additional_params
650 };
651
652 Ok(Self {
653 model: model.to_string(),
654 messages: full_history,
655 temperature: req.temperature,
656 tools: req
657 .tools
658 .clone()
659 .into_iter()
660 .map(openai::ToolDefinition::from)
661 .collect::<Vec<_>>(),
662 tool_choice,
663 additional_params,
664 })
665 }
666}
667
668#[derive(Clone)]
669pub struct CompletionModel<T = reqwest::Client> {
670 client: Client<T>,
671 pub model: String,
673}
674
675impl<T> CompletionModel<T> {
676 pub fn new(client: Client<T>, model: impl Into<String>) -> Self {
677 Self {
678 client,
679 model: model.into(),
680 }
681 }
682}
683
684impl<T> completion::CompletionModel for CompletionModel<T>
685where
686 T: HttpClientExt + Clone + Default + std::fmt::Debug + Send + 'static,
687{
688 type Response = openai::CompletionResponse;
689 type StreamingResponse = openai::StreamingCompletionResponse;
690 type Client = Client<T>;
691
692 fn make(client: &Self::Client, model: impl Into<String>) -> Self {
693 Self::new(client.clone(), model.into())
694 }
695
696 async fn completion(
697 &self,
698 completion_request: CompletionRequest,
699 ) -> Result<completion::CompletionResponse<openai::CompletionResponse>, CompletionError> {
700 let span = if tracing::Span::current().is_disabled() {
701 info_span!(
702 target: "rig::completions",
703 "chat",
704 gen_ai.operation.name = "chat",
705 gen_ai.provider.name = "azure.openai",
706 gen_ai.request.model = self.model,
707 gen_ai.system_instructions = &completion_request.preamble,
708 gen_ai.response.id = tracing::field::Empty,
709 gen_ai.response.model = tracing::field::Empty,
710 gen_ai.usage.output_tokens = tracing::field::Empty,
711 gen_ai.usage.input_tokens = tracing::field::Empty,
712 gen_ai.usage.cache_read.input_tokens = tracing::field::Empty,
713 )
714 } else {
715 tracing::Span::current()
716 };
717
718 let request =
719 AzureOpenAICompletionRequest::try_from((self.model.as_ref(), completion_request))?;
720
721 if enabled!(Level::TRACE) {
722 tracing::trace!(target: "rig::completions",
723 "Azure OpenAI completion request: {}",
724 serde_json::to_string_pretty(&request)?
725 );
726 }
727
728 let body = serde_json::to_vec(&request)?;
729
730 let req = self
731 .client
732 .post_chat_completion(&self.model)?
733 .body(body)
734 .map_err(http_client::Error::from)?;
735
736 async move {
737 let response = self.client.send::<_, Bytes>(req).await?;
738
739 let status = response.status();
740 let response_body = response.into_body().into_future().await?.to_vec();
741
742 if status.is_success() {
743 match serde_json::from_slice::<ApiResponse<openai::CompletionResponse>>(
744 &response_body,
745 )? {
746 ApiResponse::Ok(response) => {
747 let span = tracing::Span::current();
748 span.record_response_metadata(&response);
749 span.record_token_usage(&response.usage);
750 if enabled!(Level::TRACE) {
751 tracing::trace!(target: "rig::completions",
752 "Azure OpenAI completion response: {}",
753 serde_json::to_string_pretty(&response)?
754 );
755 }
756 response.try_into()
757 }
758 ApiResponse::Err(err) => Err(CompletionError::ProviderError(err.message)),
759 }
760 } else {
761 Err(CompletionError::ProviderError(
762 String::from_utf8_lossy(&response_body).to_string(),
763 ))
764 }
765 }
766 .instrument(span)
767 .await
768 }
769
770 async fn stream(
771 &self,
772 completion_request: CompletionRequest,
773 ) -> Result<StreamingCompletionResponse<Self::StreamingResponse>, CompletionError> {
774 let preamble = completion_request.preamble.clone();
775 let mut request =
776 AzureOpenAICompletionRequest::try_from((self.model.as_ref(), completion_request))?;
777
778 let params = json_utils::merge(
779 request.additional_params.unwrap_or(serde_json::json!({})),
780 serde_json::json!({"stream": true, "stream_options": {"include_usage": true} }),
781 );
782
783 request.additional_params = Some(params);
784
785 if enabled!(Level::TRACE) {
786 tracing::trace!(target: "rig::completions",
787 "Azure OpenAI completion request: {}",
788 serde_json::to_string_pretty(&request)?
789 );
790 }
791
792 let body = serde_json::to_vec(&request)?;
793
794 let req = self
795 .client
796 .post_chat_completion(&self.model)?
797 .body(body)
798 .map_err(http_client::Error::from)?;
799
800 let span = if tracing::Span::current().is_disabled() {
801 info_span!(
802 target: "rig::completions",
803 "chat_streaming",
804 gen_ai.operation.name = "chat_streaming",
805 gen_ai.provider.name = "azure.openai",
806 gen_ai.request.model = self.model,
807 gen_ai.system_instructions = &preamble,
808 gen_ai.response.id = tracing::field::Empty,
809 gen_ai.response.model = tracing::field::Empty,
810 gen_ai.usage.output_tokens = tracing::field::Empty,
811 gen_ai.usage.input_tokens = tracing::field::Empty,
812 gen_ai.usage.cache_read.input_tokens = tracing::field::Empty,
813 )
814 } else {
815 tracing::Span::current()
816 };
817
818 tracing_futures::Instrument::instrument(
819 send_compatible_streaming_request(self.client.clone(), req),
820 span,
821 )
822 .await
823 }
824}
825
826#[derive(Clone)]
831pub struct TranscriptionModel<T = reqwest::Client> {
832 client: Client<T>,
833 pub model: String,
835}
836
837impl<T> TranscriptionModel<T> {
838 pub fn new(client: Client<T>, model: impl Into<String>) -> Self {
839 Self {
840 client,
841 model: model.into(),
842 }
843 }
844}
845
846impl<T> transcription::TranscriptionModel for TranscriptionModel<T>
847where
848 T: HttpClientExt + Clone + 'static,
849{
850 type Response = TranscriptionResponse;
851 type Client = Client<T>;
852
853 fn make(client: &Self::Client, model: impl Into<String>) -> Self {
854 Self::new(client.clone(), model)
855 }
856
857 async fn transcription(
858 &self,
859 request: transcription::TranscriptionRequest,
860 ) -> Result<
861 transcription::TranscriptionResponse<Self::Response>,
862 transcription::TranscriptionError,
863 > {
864 let data = request.data;
865
866 let mut body =
867 MultipartForm::new().part(Part::bytes("file", data).filename(request.filename.clone()));
868
869 if let Some(prompt) = request.prompt {
870 body = body.text("prompt", prompt.clone());
871 }
872
873 if let Some(ref temperature) = request.temperature {
874 body = body.text("temperature", temperature.to_string());
875 }
876
877 if let Some(ref additional_params) = request.additional_params {
878 let params = additional_params.as_object().ok_or_else(|| {
879 TranscriptionError::RequestError(Box::new(std::io::Error::new(
880 std::io::ErrorKind::InvalidInput,
881 "additional transcription parameters must be a JSON object",
882 )))
883 })?;
884
885 for (key, value) in params {
886 body = body.text(key.to_owned(), value.to_string());
887 }
888 }
889
890 let req = self
891 .client
892 .post_transcription(&self.model)?
893 .body(body)
894 .map_err(|e| TranscriptionError::HttpError(e.into()))?;
895
896 let response = self.client.send_multipart::<Bytes>(req).await?;
897 let status = response.status();
898 let response_body = response.into_body().into_future().await?.to_vec();
899
900 if status.is_success() {
901 match serde_json::from_slice::<ApiResponse<TranscriptionResponse>>(&response_body)? {
902 ApiResponse::Ok(response) => response.try_into(),
903 ApiResponse::Err(api_error_response) => Err(TranscriptionError::ProviderError(
904 api_error_response.message,
905 )),
906 }
907 } else {
908 Err(TranscriptionError::ProviderError(
909 String::from_utf8_lossy(&response_body).to_string(),
910 ))
911 }
912 }
913}
914
915#[cfg(feature = "image")]
919pub use image_generation::*;
920use tracing::{Instrument, Level, enabled, info_span};
921#[cfg(feature = "image")]
922#[cfg_attr(docsrs, doc(cfg(feature = "image")))]
923mod image_generation {
924 use crate::http_client::HttpClientExt;
925 use crate::image_generation;
926 use crate::image_generation::{ImageGenerationError, ImageGenerationRequest};
927 use crate::providers::azure::{ApiResponse, Client};
928 use crate::providers::openai::ImageGenerationResponse;
929 use bytes::Bytes;
930 use serde_json::json;
931
932 #[derive(Clone)]
933 pub struct ImageGenerationModel<T = reqwest::Client> {
934 client: Client<T>,
935 pub model: String,
936 }
937
938 impl<T> image_generation::ImageGenerationModel for ImageGenerationModel<T>
939 where
940 T: HttpClientExt + Clone + Default + std::fmt::Debug + Send + 'static,
941 {
942 type Response = ImageGenerationResponse;
943
944 type Client = Client<T>;
945
946 fn make(client: &Self::Client, model: impl Into<String>) -> Self {
947 Self {
948 client: client.clone(),
949 model: model.into(),
950 }
951 }
952
953 async fn image_generation(
954 &self,
955 generation_request: ImageGenerationRequest,
956 ) -> Result<image_generation::ImageGenerationResponse<Self::Response>, ImageGenerationError>
957 {
958 let request = json!({
959 "model": self.model,
960 "prompt": generation_request.prompt,
961 "size": format!("{}x{}", generation_request.width, generation_request.height),
962 "response_format": "b64_json"
963 });
964
965 let body = serde_json::to_vec(&request)?;
966
967 let req = self
968 .client
969 .post_image_generation(&self.model)?
970 .body(body)
971 .map_err(|e| ImageGenerationError::HttpError(e.into()))?;
972
973 let response = self.client.send::<_, Bytes>(req).await?;
974 let status = response.status();
975 let response_body = response.into_body().into_future().await?.to_vec();
976
977 if !status.is_success() {
978 return Err(ImageGenerationError::ProviderError(format!(
979 "{status}: {}",
980 String::from_utf8_lossy(&response_body)
981 )));
982 }
983
984 match serde_json::from_slice::<ApiResponse<ImageGenerationResponse>>(&response_body)? {
985 ApiResponse::Ok(response) => response.try_into(),
986 ApiResponse::Err(err) => Err(ImageGenerationError::ProviderError(err.message)),
987 }
988 }
989 }
990}
991#[cfg(feature = "audio")]
996pub use audio_generation::*;
997
998#[cfg(feature = "audio")]
999#[cfg_attr(docsrs, doc(cfg(feature = "audio")))]
1000mod audio_generation {
1001 use super::Client;
1002 use crate::audio_generation::{
1003 self, AudioGenerationError, AudioGenerationRequest, AudioGenerationResponse,
1004 };
1005 use crate::http_client::HttpClientExt;
1006 use bytes::Bytes;
1007 use serde_json::json;
1008
1009 #[derive(Clone)]
1010 pub struct AudioGenerationModel<T = reqwest::Client> {
1011 client: Client<T>,
1012 model: String,
1013 }
1014
1015 impl<T> AudioGenerationModel<T> {
1016 pub fn new(client: Client<T>, deployment_name: impl Into<String>) -> Self {
1017 Self {
1018 client,
1019 model: deployment_name.into(),
1020 }
1021 }
1022 }
1023
1024 impl<T> audio_generation::AudioGenerationModel for AudioGenerationModel<T>
1025 where
1026 T: HttpClientExt + Clone + Default + std::fmt::Debug + Send + 'static,
1027 {
1028 type Response = Bytes;
1029 type Client = Client<T>;
1030
1031 fn make(client: &Self::Client, model: impl Into<String>) -> Self {
1032 Self::new(client.clone(), model)
1033 }
1034
1035 async fn audio_generation(
1036 &self,
1037 request: AudioGenerationRequest,
1038 ) -> Result<AudioGenerationResponse<Self::Response>, AudioGenerationError> {
1039 let request = json!({
1040 "model": self.model,
1041 "input": request.text,
1042 "voice": request.voice,
1043 "speed": request.speed,
1044 });
1045
1046 let body = serde_json::to_vec(&request)?;
1047
1048 let req = self
1049 .client
1050 .post_audio_generation("/audio/speech")?
1051 .header("Content-Type", "application/json")
1052 .body(body)
1053 .map_err(|e| AudioGenerationError::HttpError(e.into()))?;
1054
1055 let response = self.client.send::<_, Bytes>(req).await?;
1056 let status = response.status();
1057 let response_body = response.into_body().into_future().await?;
1058
1059 if !status.is_success() {
1060 return Err(AudioGenerationError::ProviderError(format!(
1061 "{status}: {}",
1062 String::from_utf8_lossy(&response_body)
1063 )));
1064 }
1065
1066 Ok(AudioGenerationResponse {
1067 audio: response_body.to_vec(),
1068 response: response_body,
1069 })
1070 }
1071 }
1072}
1073
1074#[cfg(test)]
1075mod azure_tests {
1076 use schemars::JsonSchema;
1077
1078 use super::*;
1079
1080 use crate::OneOrMany;
1081 use crate::client::{completion::CompletionClient, embeddings::EmbeddingsClient};
1082 use crate::completion::CompletionModel;
1083 use crate::embeddings::EmbeddingModel;
1084 use crate::prelude::TypedPrompt;
1085 use crate::providers::openai::GPT_5_MINI;
1086
1087 #[tokio::test]
1088 #[ignore]
1089 async fn test_azure_embedding() -> anyhow::Result<()> {
1090 let _ = tracing_subscriber::fmt::try_init();
1091
1092 let client = Client::from_env()?;
1093 let model = client.embedding_model(TEXT_EMBEDDING_3_SMALL);
1094 let embeddings = model.embed_texts(vec!["Hello, world!".to_string()]).await?;
1095
1096 tracing::info!("Azure embedding: {:?}", embeddings);
1097 Ok(())
1098 }
1099
1100 #[tokio::test]
1101 #[ignore]
1102 async fn test_azure_embedding_dimensions() -> anyhow::Result<()> {
1103 let _ = tracing_subscriber::fmt::try_init();
1104
1105 let ndims = 256;
1106 let client = Client::from_env()?;
1107 let model = client.embedding_model_with_ndims(TEXT_EMBEDDING_3_SMALL, ndims);
1108 let embedding = model.embed_text("Hello, world!").await?;
1109
1110 anyhow::ensure!(
1111 embedding.vec.len() == ndims,
1112 "expected embedding dimensions {ndims}, got {}",
1113 embedding.vec.len()
1114 );
1115
1116 tracing::info!("Azure dimensions embedding: {:?}", embedding);
1117 Ok(())
1118 }
1119
1120 #[tokio::test]
1121 #[ignore]
1122 async fn test_azure_completion() -> anyhow::Result<()> {
1123 let _ = tracing_subscriber::fmt::try_init();
1124
1125 let client = Client::from_env()?;
1126 let model = client.completion_model(GPT_4O_MINI);
1127 let completion = model
1128 .completion(CompletionRequest {
1129 model: None,
1130 preamble: Some("You are a helpful assistant.".to_string()),
1131 chat_history: OneOrMany::one("Hello!".into()),
1132 documents: vec![],
1133 max_tokens: Some(100),
1134 temperature: Some(0.0),
1135 tools: vec![],
1136 tool_choice: None,
1137 additional_params: None,
1138 output_schema: None,
1139 })
1140 .await?;
1141
1142 tracing::info!("Azure completion: {:?}", completion);
1143 Ok(())
1144 }
1145
1146 #[tokio::test]
1147 #[ignore]
1148 async fn test_azure_structured_output() -> anyhow::Result<()> {
1149 let _ = tracing_subscriber::fmt::try_init();
1150
1151 #[derive(Debug, Deserialize, JsonSchema)]
1152 struct Person {
1153 name: String,
1154 age: u32,
1155 }
1156
1157 let client = Client::from_env()?;
1158 let agent = client
1159 .agent(GPT_5_MINI)
1160 .preamble("You are a helpful assistant that extracts personal details.")
1161 .max_tokens(100)
1162 .output_schema::<Person>()
1163 .build();
1164
1165 let result: Person = agent
1166 .prompt_typed("Hello! My name is John Doe and I'm 54 years old.")
1167 .await?;
1168
1169 anyhow::ensure!(
1170 result.name == "John Doe",
1171 "expected name John Doe, got {}",
1172 result.name
1173 );
1174 anyhow::ensure!(result.age == 54, "expected age 54, got {}", result.age);
1175
1176 tracing::info!("Extracted person: {:?}", result);
1177 Ok(())
1178 }
1179
1180 #[tokio::test]
1181 async fn test_client_initialization() {
1182 let _client = crate::providers::azure::Client::builder()
1183 .api_key("test")
1184 .azure_endpoint("test".to_string()) .build()
1186 .expect("Client::builder() failed");
1187 }
1188}