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, RawStreamingToolCall};
9use crate::telemetry::SpanCombinator;
10use crate::{json_utils, streaming};
11use async_stream::stream;
12use futures::StreamExt;
13use serde::{Deserialize, Serialize};
14use tracing::{Level, enabled, 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            )
115        } else {
116            tracing::Span::current()
117        };
118
119        let params = json_utils::merge(
120            request.additional_params.unwrap_or(serde_json::json!({})),
121            serde_json::json!({"stream": true}),
122        );
123
124        request.additional_params = Some(params);
125
126        if enabled!(Level::TRACE) {
127            tracing::trace!(
128                target: "rig::streaming",
129                "Cohere streaming completion input: {}",
130                serde_json::to_string_pretty(&request)?
131            );
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.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                                    RawStreamingToolCall::new(tc.0, tc.1, args)
236                                ));
237
238                                current_tool_call = None;
239                            },
240
241                            _ => {}
242                        }
243                    },
244                    Err(crate::http_client::Error::StreamEnded) => {
245                        break;
246                    }
247                    Err(err) => {
248                        tracing::error!(?err, "SSE error");
249                        yield Err(CompletionError::ProviderError(err.to_string()));
250                        break;
251                    }
252                }
253            }
254
255            // Ensure event source is closed when stream ends
256            event_source.close();
257
258            yield Ok(RawStreamingChoice::FinalResponse(StreamingCompletionResponse {
259                usage: final_usage.unwrap_or_default()
260            }))
261        }.instrument(span);
262
263        Ok(streaming::StreamingCompletionResponse::stream(Box::pin(
264            stream,
265        )))
266    }
267}
268
269#[cfg(test)]
270mod tests {
271    use super::*;
272    use serde_json::json;
273
274    #[test]
275    fn test_message_content_delta_deserialization() {
276        let json = json!({
277            "type": "content-delta",
278            "delta": {
279                "message": {
280                    "content": {
281                        "text": "Hello world"
282                    }
283                }
284            }
285        });
286
287        let event: StreamingEvent = serde_json::from_value(json).unwrap();
288        match event {
289            StreamingEvent::ContentDelta { delta } => {
290                assert!(delta.is_some());
291                let message = delta.unwrap().message.unwrap();
292                let content = message.content.unwrap();
293                assert_eq!(content.text, Some("Hello world".to_string()));
294            }
295            _ => panic!("Expected ContentDelta"),
296        }
297    }
298
299    #[test]
300    fn test_tool_call_start_deserialization() {
301        let json = json!({
302            "type": "tool-call-start",
303            "delta": {
304                "message": {
305                    "tool_calls": {
306                        "id": "call_123",
307                        "function": {
308                            "name": "get_weather",
309                            "arguments": "{"
310                        }
311                    }
312                }
313            }
314        });
315
316        let event: StreamingEvent = serde_json::from_value(json).unwrap();
317        match event {
318            StreamingEvent::ToolCallStart { delta } => {
319                assert!(delta.is_some());
320                let tool_call = delta.unwrap().message.unwrap().tool_calls.unwrap();
321                assert_eq!(tool_call.id, Some("call_123".to_string()));
322                assert_eq!(
323                    tool_call.function.unwrap().name,
324                    Some("get_weather".to_string())
325                );
326            }
327            _ => panic!("Expected ToolCallStart"),
328        }
329    }
330
331    #[test]
332    fn test_tool_call_delta_deserialization() {
333        let json = json!({
334            "type": "tool-call-delta",
335            "delta": {
336                "message": {
337                    "tool_calls": {
338                        "function": {
339                            "arguments": "\"location\""
340                        }
341                    }
342                }
343            }
344        });
345
346        let event: StreamingEvent = serde_json::from_value(json).unwrap();
347        match event {
348            StreamingEvent::ToolCallDelta { delta } => {
349                assert!(delta.is_some());
350                let tool_call = delta.unwrap().message.unwrap().tool_calls.unwrap();
351                let function = tool_call.function.unwrap();
352                assert_eq!(function.arguments, Some("\"location\"".to_string()));
353            }
354            _ => panic!("Expected ToolCallDelta"),
355        }
356    }
357
358    #[test]
359    fn test_tool_call_end_deserialization() {
360        let json = json!({
361            "type": "tool-call-end"
362        });
363
364        let event: StreamingEvent = serde_json::from_value(json).unwrap();
365        match event {
366            StreamingEvent::ToolCallEnd => {
367                // Success
368            }
369            _ => panic!("Expected ToolCallEnd"),
370        }
371    }
372
373    #[test]
374    fn test_message_end_with_usage_deserialization() {
375        let json = json!({
376            "type": "message-end",
377            "delta": {
378                "usage": {
379                    "tokens": {
380                        "input_tokens": 100,
381                        "output_tokens": 50
382                    }
383                }
384            }
385        });
386
387        let event: StreamingEvent = serde_json::from_value(json).unwrap();
388        match event {
389            StreamingEvent::MessageEnd { delta } => {
390                assert!(delta.is_some());
391                let usage = delta.unwrap().usage.unwrap();
392                let tokens = usage.tokens.unwrap();
393                assert_eq!(tokens.input_tokens, Some(100.0));
394                assert_eq!(tokens.output_tokens, Some(50.0));
395            }
396            _ => panic!("Expected MessageEnd"),
397        }
398    }
399
400    #[test]
401    fn test_streaming_event_order() {
402        // Test that a typical sequence of events deserializes correctly
403        let events = vec![
404            json!({"type": "message-start"}),
405            json!({"type": "content-start"}),
406            json!({
407                "type": "content-delta",
408                "delta": {
409                    "message": {
410                        "content": {
411                            "text": "Sure, "
412                        }
413                    }
414                }
415            }),
416            json!({
417                "type": "content-delta",
418                "delta": {
419                    "message": {
420                        "content": {
421                            "text": "I can help with that."
422                        }
423                    }
424                }
425            }),
426            json!({"type": "content-end"}),
427            json!({"type": "tool-plan"}),
428            json!({
429                "type": "tool-call-start",
430                "delta": {
431                    "message": {
432                        "tool_calls": {
433                            "id": "call_abc",
434                            "function": {
435                                "name": "search",
436                                "arguments": ""
437                            }
438                        }
439                    }
440                }
441            }),
442            json!({
443                "type": "tool-call-delta",
444                "delta": {
445                    "message": {
446                        "tool_calls": {
447                            "function": {
448                                "arguments": "{\"query\":"
449                            }
450                        }
451                    }
452                }
453            }),
454            json!({
455                "type": "tool-call-delta",
456                "delta": {
457                    "message": {
458                        "tool_calls": {
459                            "function": {
460                                "arguments": "\"Rust\"}"
461                            }
462                        }
463                    }
464                }
465            }),
466            json!({"type": "tool-call-end"}),
467            json!({
468                "type": "message-end",
469                "delta": {
470                    "usage": {
471                        "tokens": {
472                            "input_tokens": 50,
473                            "output_tokens": 25
474                        }
475                    }
476                }
477            }),
478        ];
479
480        for (i, event_json) in events.iter().enumerate() {
481            let result = serde_json::from_value::<StreamingEvent>(event_json.clone());
482            assert!(
483                result.is_ok(),
484                "Failed to deserialize event at index {}: {:?}",
485                i,
486                result.err()
487            );
488        }
489    }
490}