rig/providers/xai/
completion.rs

1//! xAI Completion Integration
2//!
3//! Uses the xAI Responses API: <https://docs.x.ai/docs/guides/chat>
4
5use bytes::Bytes;
6use serde::{Deserialize, Serialize};
7use serde_json::Value;
8use tracing::{Instrument, Level, enabled, info_span};
9
10use super::api::{ApiResponse, Message, ToolDefinition};
11use super::client::Client;
12use crate::OneOrMany;
13use crate::completion::{self, CompletionError, CompletionRequest};
14use crate::http_client::HttpClientExt;
15use crate::providers::openai::completion::ToolChoice;
16use crate::providers::openai::responses_api::streaming::StreamingCompletionResponse;
17use crate::providers::openai::responses_api::{Output, ResponsesUsage};
18use crate::streaming::StreamingCompletionResponse as BaseStreamingCompletionResponse;
19
20/// xAI completion models as of 2025-06-04
21pub const GROK_2_1212: &str = "grok-2-1212";
22pub const GROK_2_VISION_1212: &str = "grok-2-vision-1212";
23pub const GROK_3: &str = "grok-3";
24pub const GROK_3_FAST: &str = "grok-3-fast";
25pub const GROK_3_MINI: &str = "grok-3-mini";
26pub const GROK_3_MINI_FAST: &str = "grok-3-mini-fast";
27pub const GROK_2_IMAGE_1212: &str = "grok-2-image-1212";
28pub const GROK_4: &str = "grok-4-0709";
29
30// ================================================================
31// Request Types
32// ================================================================
33
34#[derive(Debug, Serialize, Deserialize)]
35pub(super) struct XAICompletionRequest {
36    model: String,
37    pub input: Vec<Message>,
38    #[serde(skip_serializing_if = "Option::is_none")]
39    temperature: Option<f64>,
40    #[serde(skip_serializing_if = "Option::is_none")]
41    max_output_tokens: Option<u64>,
42    #[serde(skip_serializing_if = "Vec::is_empty")]
43    tools: Vec<Value>,
44    #[serde(skip_serializing_if = "Option::is_none")]
45    tool_choice: Option<ToolChoice>,
46    #[serde(flatten, skip_serializing_if = "Option::is_none")]
47    pub additional_params: Option<serde_json::Value>,
48}
49
50impl TryFrom<(&str, CompletionRequest)> for XAICompletionRequest {
51    type Error = CompletionError;
52
53    fn try_from((model, req): (&str, CompletionRequest)) -> Result<Self, Self::Error> {
54        if req.output_schema.is_some() {
55            tracing::warn!("Structured outputs currently not supported for xAI");
56        }
57        let model = req.model.clone().unwrap_or_else(|| model.to_string());
58        let mut input: Vec<Message> = req
59            .preamble
60            .as_ref()
61            .map_or_else(Vec::new, |p| vec![Message::system(p)]);
62
63        if let Some(docs) = req.normalized_documents() {
64            let docs: Vec<Message> = docs.try_into()?;
65            input.extend(docs);
66        }
67
68        let mut additional_params_payload = req.additional_params.unwrap_or(Value::Null);
69
70        for msg in req.chat_history {
71            let msg: Vec<Message> = msg.try_into()?;
72            input.extend(msg);
73        }
74
75        let tool_choice = req.tool_choice.map(ToolChoice::try_from).transpose()?;
76        let mut additional_tools =
77            extract_tools_from_additional_params(&mut additional_params_payload)?;
78        let mut tools = req
79            .tools
80            .into_iter()
81            .map(ToolDefinition::from)
82            .map(serde_json::to_value)
83            .collect::<Result<Vec<_>, _>>()?;
84        tools.append(&mut additional_tools);
85        let additional_params = if additional_params_payload.is_null() {
86            None
87        } else {
88            Some(additional_params_payload)
89        };
90
91        Ok(Self {
92            model: model.to_string(),
93            input,
94            temperature: req.temperature,
95            max_output_tokens: req.max_tokens,
96            tools,
97            tool_choice,
98            additional_params,
99        })
100    }
101}
102
103fn extract_tools_from_additional_params(
104    additional_params: &mut Value,
105) -> Result<Vec<Value>, CompletionError> {
106    if let Some(map) = additional_params.as_object_mut()
107        && let Some(raw_tools) = map.remove("tools")
108    {
109        return serde_json::from_value::<Vec<Value>>(raw_tools).map_err(|err| {
110            CompletionError::RequestError(
111                format!("Invalid xAI `additional_params.tools` payload: {err}").into(),
112            )
113        });
114    }
115
116    Ok(Vec::new())
117}
118
119// ================================================================
120// Response Types
121// ================================================================
122
123#[derive(Debug, Deserialize, Serialize)]
124pub struct CompletionResponse {
125    pub id: String,
126    pub model: String,
127    pub output: Vec<Output>,
128    #[serde(default)]
129    pub created: i64,
130    #[serde(default)]
131    pub object: String,
132    #[serde(default)]
133    pub status: Option<String>,
134    pub usage: Option<ResponsesUsage>,
135}
136
137impl TryFrom<CompletionResponse> for completion::CompletionResponse<CompletionResponse> {
138    type Error = CompletionError;
139
140    fn try_from(response: CompletionResponse) -> Result<Self, Self::Error> {
141        let content: Vec<completion::AssistantContent> = response
142            .output
143            .iter()
144            .cloned()
145            .flat_map(<Vec<completion::AssistantContent>>::from)
146            .collect();
147
148        let choice = OneOrMany::many(content).map_err(|_| {
149            CompletionError::ResponseError("Response contained no output".to_owned())
150        })?;
151
152        let usage = response
153            .usage
154            .as_ref()
155            .map(|u| completion::Usage {
156                input_tokens: u.input_tokens,
157                output_tokens: u.output_tokens,
158                total_tokens: u.total_tokens,
159                cached_input_tokens: u
160                    .input_tokens_details
161                    .clone()
162                    .map(|x| x.cached_tokens)
163                    .unwrap_or_default(),
164                cache_creation_input_tokens: 0,
165            })
166            .unwrap_or_default();
167
168        Ok(completion::CompletionResponse {
169            choice,
170            usage,
171            raw_response: response,
172            message_id: None,
173        })
174    }
175}
176
177// ================================================================
178// Completion Model
179// ================================================================
180
181#[derive(Clone)]
182pub struct CompletionModel<T = reqwest::Client> {
183    pub(crate) client: Client<T>,
184    pub model: String,
185}
186
187impl<T> CompletionModel<T> {
188    pub fn new(client: Client<T>, model: impl Into<String>) -> Self {
189        Self {
190            client,
191            model: model.into(),
192        }
193    }
194}
195
196impl<T> completion::CompletionModel for CompletionModel<T>
197where
198    T: HttpClientExt + Clone + Default + std::fmt::Debug + Send + 'static,
199{
200    type Response = CompletionResponse;
201    type StreamingResponse = StreamingCompletionResponse;
202
203    type Client = Client<T>;
204
205    fn make(client: &Self::Client, model: impl Into<String>) -> Self {
206        Self::new(client.clone(), model)
207    }
208
209    async fn completion(
210        &self,
211        completion_request: completion::CompletionRequest,
212    ) -> Result<completion::CompletionResponse<CompletionResponse>, CompletionError> {
213        let span = if tracing::Span::current().is_disabled() {
214            info_span!(
215                target: "rig::completions",
216                "chat",
217                gen_ai.operation.name = "chat",
218                gen_ai.provider.name = "xai",
219                gen_ai.request.model = self.model,
220                gen_ai.system_instructions = tracing::field::Empty,
221                gen_ai.response.id = tracing::field::Empty,
222                gen_ai.response.model = tracing::field::Empty,
223                gen_ai.usage.output_tokens = tracing::field::Empty,
224                gen_ai.usage.input_tokens = tracing::field::Empty,
225                gen_ai.usage.cache_read.input_tokens = tracing::field::Empty,
226            )
227        } else {
228            tracing::Span::current()
229        };
230
231        span.record("gen_ai.system_instructions", &completion_request.preamble);
232
233        let request =
234            XAICompletionRequest::try_from((self.model.to_string().as_ref(), completion_request))?;
235
236        if enabled!(Level::TRACE) {
237            tracing::trace!(target: "rig::completions",
238                "xAI completion request: {}",
239                serde_json::to_string_pretty(&request)?
240            );
241        }
242
243        let body = serde_json::to_vec(&request)?;
244        let req = self
245            .client
246            .post("/v1/responses")?
247            .body(body)
248            .map_err(|e| CompletionError::HttpError(e.into()))?;
249
250        async move {
251            let response = self.client.send::<_, Bytes>(req).await?;
252            let status = response.status();
253            let response_body = response.into_body().into_future().await?.to_vec();
254
255            if status.is_success() {
256                match serde_json::from_slice::<ApiResponse<CompletionResponse>>(&response_body)? {
257                    ApiResponse::Ok(response) => {
258                        if enabled!(Level::TRACE) {
259                            tracing::trace!(target: "rig::completions",
260                                "xAI completion response: {}",
261                                serde_json::to_string_pretty(&response)?
262                            );
263                        }
264
265                        response.try_into()
266                    }
267                    ApiResponse::Error(error) => {
268                        Err(CompletionError::ProviderError(error.message()))
269                    }
270                }
271            } else {
272                Err(CompletionError::ProviderError(
273                    String::from_utf8_lossy(&response_body).to_string(),
274                ))
275            }
276        }
277        .instrument(span)
278        .await
279    }
280
281    async fn stream(
282        &self,
283        request: CompletionRequest,
284    ) -> Result<BaseStreamingCompletionResponse<Self::StreamingResponse>, CompletionError> {
285        self.stream(request).await
286    }
287}
288
289#[cfg(test)]
290mod tests {
291    use super::XAICompletionRequest;
292    use crate::OneOrMany;
293    use crate::completion::CompletionRequest;
294    use crate::completion::request::Document;
295
296    #[test]
297    fn xai_request_includes_normalized_documents() {
298        let request = CompletionRequest {
299            model: None,
300            preamble: Some("Use the provided context.".to_string()),
301            chat_history: OneOrMany::one("What is glarb-glarb?".into()),
302            documents: vec![Document {
303                id: "doc_1".to_string(),
304                text: "Definition of glarb-glarb: an ancient tool.".to_string(),
305                additional_props: Default::default(),
306            }],
307            tools: vec![],
308            temperature: None,
309            max_tokens: None,
310            tool_choice: None,
311            additional_params: None,
312            output_schema: None,
313        };
314
315        let xai_request = XAICompletionRequest::try_from(("grok-4-0709", request))
316            .expect("request conversion should succeed");
317        let serialized = serde_json::to_value(xai_request).expect("serialization should succeed");
318        let input = serialized["input"]
319            .as_array()
320            .expect("xAI request input should be an array");
321
322        assert!(
323            input
324                .iter()
325                .any(|message| message.to_string().contains("glarb-glarb")),
326            "normalized documents should be forwarded into xAI input"
327        );
328    }
329}
rig/providers/xai/completion.rs

rig/providers/xai/
completion.rs
