rig/providers/cohere/
streaming.rs

1use crate::completion::{CompletionError, CompletionRequest, GetTokenUsage};
2use crate::http_client::HttpClientExt;
3use crate::http_client::sse::{Event, GenericEventSource};
4use crate::providers::cohere::CompletionModel;
5use crate::providers::cohere::completion::{
6    AssistantContent, CohereCompletionRequest, Message, ToolCall, ToolCallFunction, ToolType, Usage,
7};
8use crate::streaming::RawStreamingChoice;
9use crate::telemetry::SpanCombinator;
10use crate::{json_utils, streaming};
11use async_stream::stream;
12use futures::StreamExt;
13use serde::{Deserialize, Serialize};
14use tracing::info_span;
15use tracing_futures::Instrument;
16
17#[derive(Debug, Deserialize)]
18#[serde(rename_all = "kebab-case", tag = "type")]
19enum StreamingEvent {
20    MessageStart,
21    ContentStart,
22    ContentDelta { delta: Option<Delta> },
23    ContentEnd,
24    ToolPlan,
25    ToolCallStart { delta: Option<Delta> },
26    ToolCallDelta { delta: Option<Delta> },
27    ToolCallEnd,
28    MessageEnd { delta: Option<MessageEndDelta> },
29}
30
31#[derive(Debug, Deserialize)]
32struct MessageContentDelta {
33    text: Option<String>,
34}
35
36#[derive(Debug, Deserialize)]
37struct MessageToolFunctionDelta {
38    name: Option<String>,
39    arguments: Option<String>,
40}
41
42#[derive(Debug, Deserialize)]
43struct MessageToolCallDelta {
44    id: Option<String>,
45    function: Option<MessageToolFunctionDelta>,
46}
47
48#[derive(Debug, Deserialize)]
49struct MessageDelta {
50    content: Option<MessageContentDelta>,
51    tool_calls: Option<MessageToolCallDelta>,
52}
53
54#[derive(Debug, Deserialize)]
55struct Delta {
56    message: Option<MessageDelta>,
57}
58
59#[derive(Debug, Deserialize)]
60struct MessageEndDelta {
61    usage: Option<Usage>,
62}
63
64#[derive(Clone, Serialize, Deserialize)]
65pub struct StreamingCompletionResponse {
66    pub usage: Option<Usage>,
67}
68
69impl GetTokenUsage for StreamingCompletionResponse {
70    fn token_usage(&self) -> Option<crate::completion::Usage> {
71        let tokens = self
72            .usage
73            .clone()
74            .and_then(|response| response.tokens)
75            .map(|tokens| {
76                (
77                    tokens.input_tokens.map(|x| x as u64),
78                    tokens.output_tokens.map(|y| y as u64),
79                )
80            });
81        let Some((Some(input), Some(output))) = tokens else {
82            return None;
83        };
84        let mut usage = crate::completion::Usage::new();
85        usage.input_tokens = input;
86        usage.output_tokens = output;
87        usage.total_tokens = input + output;
88
89        Some(usage)
90    }
91}
92
93impl<T> CompletionModel<T>
94where
95    T: HttpClientExt + Clone + 'static,
96{
97    pub(crate) async fn stream(
98        &self,
99        request: CompletionRequest,
100    ) -> Result<streaming::StreamingCompletionResponse<StreamingCompletionResponse>, CompletionError>
101    {
102        let mut request = CohereCompletionRequest::try_from((self.model.as_ref(), request))?;
103        let span = if tracing::Span::current().is_disabled() {
104            info_span!(
105                target: "rig::completions",
106                "chat_streaming",
107                gen_ai.operation.name = "chat_streaming",
108                gen_ai.provider.name = "cohere",
109                gen_ai.request.model = self.model,
110                gen_ai.response.id = tracing::field::Empty,
111                gen_ai.response.model = self.model,
112                gen_ai.usage.output_tokens = tracing::field::Empty,
113                gen_ai.usage.input_tokens = tracing::field::Empty,
114                gen_ai.input.messages = serde_json::to_string(&request.messages)?,
115                gen_ai.output.messages = tracing::field::Empty,
116            )
117        } else {
118            tracing::Span::current()
119        };
120
121        let params = json_utils::merge(
122            request.additional_params.unwrap_or(serde_json::json!({})),
123            serde_json::json!({"stream": true}),
124        );
125
126        request.additional_params = Some(params);
127
128        tracing::trace!(
129            target: "rig::streaming",
130            "Cohere streaming completion input: {}",
131            serde_json::to_string_pretty(&request)?
132        );
133
134        let body = serde_json::to_vec(&request)?;
135
136        let req = self.client.post("/v2/chat")?.body(body).unwrap();
137
138        let mut event_source = GenericEventSource::new(self.client.http_client().clone(), req);
139
140        let stream = stream! {
141            let mut current_tool_call: Option<(String, String, String)> = None;
142            let mut text_response = String::new();
143            let mut tool_calls = Vec::new();
144            let mut final_usage = None;
145
146            while let Some(event_result) = event_source.next().await {
147                match event_result {
148                    Ok(Event::Open) => {
149                        tracing::trace!("SSE connection opened");
150                        continue;
151                    }
152
153                    Ok(Event::Message(message)) => {
154                        let data_str = message.data.trim();
155                        if data_str.is_empty() || data_str == "[DONE]" {
156                            continue;
157                        }
158
159                        let event: StreamingEvent = match serde_json::from_str(data_str) {
160                            Ok(ev) => ev,
161                            Err(_) => {
162                                tracing::debug!("Couldn't parse SSE payload as StreamingEvent");
163                                continue;
164                            }
165                        };
166
167                        match event {
168                            StreamingEvent::ContentDelta { delta: Some(delta) } => {
169                                let Some(message) = &delta.message else { continue; };
170                                let Some(content) = &message.content else { continue; };
171                                let Some(text) = &content.text else { continue; };
172
173                                text_response += text;
174
175                                yield Ok(RawStreamingChoice::Message(text.clone()));
176                            },
177
178                            StreamingEvent::MessageEnd { delta: Some(delta) } => {
179                                let message = Message::Assistant {
180                                    tool_calls: tool_calls.clone(),
181                                    content: vec![AssistantContent::Text { text: text_response.clone() }],
182                                    tool_plan: None,
183                                    citations: vec![]
184                                };
185
186                                let span = tracing::Span::current();
187                                span.record_token_usage(&delta.usage);
188                                span.record_model_output(&vec![message]);
189
190                                final_usage = Some(delta.usage.clone());
191                                break;
192                            },
193
194                            StreamingEvent::ToolCallStart { delta: Some(delta) } => {
195                                let Some(message) = &delta.message else { continue; };
196                                let Some(tool_calls) = &message.tool_calls else { continue; };
197                                let Some(id) = tool_calls.id.clone() else { continue; };
198                                let Some(function) = &tool_calls.function else { continue; };
199                                let Some(name) = function.name.clone() else { continue; };
200                                let Some(arguments) = function.arguments.clone() else { continue; };
201
202                                current_tool_call = Some((id, name, arguments));
203                            },
204
205                            StreamingEvent::ToolCallDelta { delta: Some(delta) } => {
206                                let Some(message) = &delta.message else { continue; };
207                                let Some(tool_calls) = &message.tool_calls else { continue; };
208                                let Some(function) = &tool_calls.function else { continue; };
209                                let Some(arguments) = function.arguments.clone() else { continue; };
210
211                                let Some(tc) = current_tool_call.clone() else { continue; };
212                                current_tool_call = Some((tc.0.clone(), tc.1, format!("{}{}", tc.2, arguments)));
213
214                                // Emit the delta so UI can show progress
215                                yield Ok(RawStreamingChoice::ToolCallDelta {
216                                    id: tc.0,
217                                    delta: arguments,
218                                });
219                            },
220
221                            StreamingEvent::ToolCallEnd => {
222                                let Some(tc) = current_tool_call.clone() else { continue; };
223                                let Ok(args) = serde_json::from_str::<serde_json::Value>(&tc.2) else { continue; };
224
225                                tool_calls.push(ToolCall {
226                                    id: Some(tc.0.clone()),
227                                    r#type: Some(ToolType::Function),
228                                    function: Some(ToolCallFunction {
229                                        name: tc.1.clone(),
230                                        arguments: args.clone()
231                                    })
232                                });
233
234                                yield Ok(RawStreamingChoice::ToolCall {
235                                    id: tc.0,
236                                    name: tc.1,
237                                    arguments: args,
238                                    call_id: None
239                                });
240
241                                current_tool_call = None;
242                            },
243
244                            _ => {}
245                        }
246                    },
247                    Err(crate::http_client::Error::StreamEnded) => {
248                        break;
249                    }
250                    Err(err) => {
251                        tracing::error!(?err, "SSE error");
252                        yield Err(CompletionError::ProviderError(err.to_string()));
253                        break;
254                    }
255                }
256            }
257
258            // Ensure event source is closed when stream ends
259            event_source.close();
260
261            yield Ok(RawStreamingChoice::FinalResponse(StreamingCompletionResponse {
262                usage: final_usage.unwrap_or_default()
263            }))
264        }.instrument(span);
265
266        Ok(streaming::StreamingCompletionResponse::stream(Box::pin(
267            stream,
268        )))
269    }
270}
271
272#[cfg(test)]
273mod tests {
274    use super::*;
275    use serde_json::json;
276
277    #[test]
278    fn test_message_content_delta_deserialization() {
279        let json = json!({
280            "type": "content-delta",
281            "delta": {
282                "message": {
283                    "content": {
284                        "text": "Hello world"
285                    }
286                }
287            }
288        });
289
290        let event: StreamingEvent = serde_json::from_value(json).unwrap();
291        match event {
292            StreamingEvent::ContentDelta { delta } => {
293                assert!(delta.is_some());
294                let message = delta.unwrap().message.unwrap();
295                let content = message.content.unwrap();
296                assert_eq!(content.text, Some("Hello world".to_string()));
297            }
298            _ => panic!("Expected ContentDelta"),
299        }
300    }
301
302    #[test]
303    fn test_tool_call_start_deserialization() {
304        let json = json!({
305            "type": "tool-call-start",
306            "delta": {
307                "message": {
308                    "tool_calls": {
309                        "id": "call_123",
310                        "function": {
311                            "name": "get_weather",
312                            "arguments": "{"
313                        }
314                    }
315                }
316            }
317        });
318
319        let event: StreamingEvent = serde_json::from_value(json).unwrap();
320        match event {
321            StreamingEvent::ToolCallStart { delta } => {
322                assert!(delta.is_some());
323                let tool_call = delta.unwrap().message.unwrap().tool_calls.unwrap();
324                assert_eq!(tool_call.id, Some("call_123".to_string()));
325                assert_eq!(
326                    tool_call.function.unwrap().name,
327                    Some("get_weather".to_string())
328                );
329            }
330            _ => panic!("Expected ToolCallStart"),
331        }
332    }
333
334    #[test]
335    fn test_tool_call_delta_deserialization() {
336        let json = json!({
337            "type": "tool-call-delta",
338            "delta": {
339                "message": {
340                    "tool_calls": {
341                        "function": {
342                            "arguments": "\"location\""
343                        }
344                    }
345                }
346            }
347        });
348
349        let event: StreamingEvent = serde_json::from_value(json).unwrap();
350        match event {
351            StreamingEvent::ToolCallDelta { delta } => {
352                assert!(delta.is_some());
353                let tool_call = delta.unwrap().message.unwrap().tool_calls.unwrap();
354                let function = tool_call.function.unwrap();
355                assert_eq!(function.arguments, Some("\"location\"".to_string()));
356            }
357            _ => panic!("Expected ToolCallDelta"),
358        }
359    }
360
361    #[test]
362    fn test_tool_call_end_deserialization() {
363        let json = json!({
364            "type": "tool-call-end"
365        });
366
367        let event: StreamingEvent = serde_json::from_value(json).unwrap();
368        match event {
369            StreamingEvent::ToolCallEnd => {
370                // Success
371            }
372            _ => panic!("Expected ToolCallEnd"),
373        }
374    }
375
376    #[test]
377    fn test_message_end_with_usage_deserialization() {
378        let json = json!({
379            "type": "message-end",
380            "delta": {
381                "usage": {
382                    "tokens": {
383                        "input_tokens": 100,
384                        "output_tokens": 50
385                    }
386                }
387            }
388        });
389
390        let event: StreamingEvent = serde_json::from_value(json).unwrap();
391        match event {
392            StreamingEvent::MessageEnd { delta } => {
393                assert!(delta.is_some());
394                let usage = delta.unwrap().usage.unwrap();
395                let tokens = usage.tokens.unwrap();
396                assert_eq!(tokens.input_tokens, Some(100.0));
397                assert_eq!(tokens.output_tokens, Some(50.0));
398            }
399            _ => panic!("Expected MessageEnd"),
400        }
401    }
402
403    #[test]
404    fn test_streaming_event_order() {
405        // Test that a typical sequence of events deserializes correctly
406        let events = vec![
407            json!({"type": "message-start"}),
408            json!({"type": "content-start"}),
409            json!({
410                "type": "content-delta",
411                "delta": {
412                    "message": {
413                        "content": {
414                            "text": "Sure, "
415                        }
416                    }
417                }
418            }),
419            json!({
420                "type": "content-delta",
421                "delta": {
422                    "message": {
423                        "content": {
424                            "text": "I can help with that."
425                        }
426                    }
427                }
428            }),
429            json!({"type": "content-end"}),
430            json!({"type": "tool-plan"}),
431            json!({
432                "type": "tool-call-start",
433                "delta": {
434                    "message": {
435                        "tool_calls": {
436                            "id": "call_abc",
437                            "function": {
438                                "name": "search",
439                                "arguments": ""
440                            }
441                        }
442                    }
443                }
444            }),
445            json!({
446                "type": "tool-call-delta",
447                "delta": {
448                    "message": {
449                        "tool_calls": {
450                            "function": {
451                                "arguments": "{\"query\":"
452                            }
453                        }
454                    }
455                }
456            }),
457            json!({
458                "type": "tool-call-delta",
459                "delta": {
460                    "message": {
461                        "tool_calls": {
462                            "function": {
463                                "arguments": "\"Rust\"}"
464                            }
465                        }
466                    }
467                }
468            }),
469            json!({"type": "tool-call-end"}),
470            json!({
471                "type": "message-end",
472                "delta": {
473                    "usage": {
474                        "tokens": {
475                            "input_tokens": 50,
476                            "output_tokens": 25
477                        }
478                    }
479                }
480            }),
481        ];
482
483        for (i, event_json) in events.iter().enumerate() {
484            let result = serde_json::from_value::<StreamingEvent>(event_json.clone());
485            assert!(
486                result.is_ok(),
487                "Failed to deserialize event at index {}: {:?}",
488                i,
489                result.err()
490            );
491        }
492    }
493}