rig/providers/together/
completion.rs

1// ================================================================
2//! Together AI Completion Integration
3//! From [Together AI Reference](https://docs.together.ai/docs/chat-overview)
4// ================================================================
5
6use crate::{
7    completion::{self, CompletionError},
8    http_client::HttpClientExt,
9    providers::openai,
10};
11
12use super::client::{Client, together_ai_api_types::ApiResponse};
13use crate::completion::CompletionRequest;
14use crate::streaming::StreamingCompletionResponse;
15use bytes::Bytes;
16use serde::{Deserialize, Serialize};
17use tracing::{Instrument, info_span};
18
19// ================================================================
20// Together Completion Models
21// ================================================================
22
23pub const YI_34B_CHAT: &str = "zero-one-ai/Yi-34B-Chat";
24pub const OLMO_7B_INSTRUCT: &str = "allenai/OLMo-7B-Instruct";
25pub const CHRONOS_HERMES_13B: &str = "Austism/chronos-hermes-13b";
26pub const ML318BR: &str = "carson/ml318br";
27pub const DOLPHIN_2_5_MIXTRAL_8X7B: &str = "cognitivecomputations/dolphin-2.5-mixtral-8x7b";
28pub const DBRX_INSTRUCT: &str = "databricks/dbrx-instruct";
29pub const DEEPSEEK_LLM_67B_CHAT: &str = "deepseek-ai/deepseek-llm-67b-chat";
30pub const DEEPSEEK_CODER_33B_INSTRUCT: &str = "deepseek-ai/deepseek-coder-33b-instruct";
31pub const PLATYPUS2_70B_INSTRUCT: &str = "garage-bAInd/Platypus2-70B-instruct";
32pub const GEMMA_2_9B_IT: &str = "google/gemma-2-9b-it";
33pub const GEMMA_2B_IT: &str = "google/gemma-2b-it";
34pub const GEMMA_2_27B_IT: &str = "google/gemma-2-27b-it";
35pub const GEMMA_7B_IT: &str = "google/gemma-7b-it";
36pub const LLAMA_3_70B_INSTRUCT_GRADIENT_1048K: &str =
37    "gradientai/Llama-3-70B-Instruct-Gradient-1048k";
38pub const MYTHOMAX_L2_13B: &str = "Gryphe/MythoMax-L2-13b";
39pub const MYTHOMAX_L2_13B_LITE: &str = "Gryphe/MythoMax-L2-13b-Lite";
40pub const LLAVA_NEXT_MISTRAL_7B: &str = "llava-hf/llava-v1.6-mistral-7b-hf";
41pub const ZEPHYR_7B_BETA: &str = "HuggingFaceH4/zephyr-7b-beta";
42pub const KOALA_7B: &str = "togethercomputer/Koala-7B";
43pub const VICUNA_7B_V1_3: &str = "lmsys/vicuna-7b-v1.3";
44pub const VICUNA_13B_V1_5_16K: &str = "lmsys/vicuna-13b-v1.5-16k";
45pub const VICUNA_13B_V1_5: &str = "lmsys/vicuna-13b-v1.5";
46pub const VICUNA_13B_V1_3: &str = "lmsys/vicuna-13b-v1.3";
47pub const KOALA_13B: &str = "togethercomputer/Koala-13B";
48pub const VICUNA_7B_V1_5: &str = "lmsys/vicuna-7b-v1.5";
49pub const CODE_LLAMA_34B_INSTRUCT: &str = "codellama/CodeLlama-34b-Instruct-hf";
50pub const LLAMA_3_8B_CHAT_HF_INT4: &str = "togethercomputer/Llama-3-8b-chat-hf-int4";
51pub const LLAMA_3_2_90B_VISION_INSTRUCT_TURBO: &str =
52    "meta-llama/Llama-3.2-90B-Vision-Instruct-Turbo";
53pub const LLAMA_3_2_11B_VISION_INSTRUCT_TURBO: &str =
54    "meta-llama/Llama-3.2-11B-Vision-Instruct-Turbo";
55pub const LLAMA_3_2_3B_INSTRUCT_TURBO: &str = "meta-llama/Llama-3.2-3B-Instruct-Turbo";
56pub const LLAMA_3_8B_CHAT_HF_INT8: &str = "togethercomputer/Llama-3-8b-chat-hf-int8";
57pub const LLAMA_3_1_70B_INSTRUCT_TURBO: &str = "meta-llama/Meta-Llama-3.1-70B-Instruct-Turbo";
58pub const LLAMA_2_13B_CHAT: &str = "meta-llama/Llama-2-13b-chat-hf";
59pub const LLAMA_3_70B_INSTRUCT_LITE: &str = "meta-llama/Meta-Llama-3-70B-Instruct-Lite";
60pub const LLAMA_3_8B_CHAT_HF: &str = "meta-llama/Llama-3-8b-chat-hf";
61pub const LLAMA_3_70B_CHAT_HF: &str = "meta-llama/Llama-3-70b-chat-hf";
62pub const LLAMA_3_8B_INSTRUCT_TURBO: &str = "meta-llama/Meta-Llama-3-8B-Instruct-Turbo";
63pub const LLAMA_3_8B_INSTRUCT_LITE: &str = "meta-llama/Meta-Llama-3-8B-Instruct-Lite";
64pub const LLAMA_3_1_405B_INSTRUCT_LITE_PRO: &str =
65    "meta-llama/Meta-Llama-3.1-405B-Instruct-Lite-Pro";
66pub const LLAMA_2_7B_CHAT: &str = "meta-llama/Llama-2-7b-chat-hf";
67pub const LLAMA_3_1_405B_INSTRUCT_TURBO: &str = "meta-llama/Meta-Llama-3.1-405B-Instruct-Turbo";
68pub const LLAMA_VISION_FREE: &str = "meta-llama/Llama-Vision-Free";
69pub const LLAMA_3_70B_INSTRUCT_TURBO: &str = "meta-llama/Meta-Llama-3-70B-Instruct-Turbo";
70pub const LLAMA_3_1_8B_INSTRUCT_TURBO: &str = "meta-llama/Meta-Llama-3.1-8B-Instruct-Turbo";
71pub const CODE_LLAMA_7B_INSTRUCT_TOGETHER: &str = "togethercomputer/CodeLlama-7b-Instruct";
72pub const CODE_LLAMA_34B_INSTRUCT_TOGETHER: &str = "togethercomputer/CodeLlama-34b-Instruct";
73pub const CODE_LLAMA_13B_INSTRUCT: &str = "codellama/CodeLlama-13b-Instruct-hf";
74pub const CODE_LLAMA_13B_INSTRUCT_TOGETHER: &str = "togethercomputer/CodeLlama-13b-Instruct";
75pub const LLAMA_2_13B_CHAT_TOGETHER: &str = "togethercomputer/llama-2-13b-chat";
76pub const LLAMA_2_7B_CHAT_TOGETHER: &str = "togethercomputer/llama-2-7b-chat";
77pub const LLAMA_3_8B_INSTRUCT: &str = "meta-llama/Meta-Llama-3-8B-Instruct";
78pub const LLAMA_3_70B_INSTRUCT: &str = "meta-llama/Meta-Llama-3-70B-Instruct";
79pub const CODE_LLAMA_70B_INSTRUCT: &str = "codellama/CodeLlama-70b-Instruct-hf";
80pub const LLAMA_2_70B_CHAT_TOGETHER: &str = "togethercomputer/llama-2-70b-chat";
81pub const LLAMA_3_1_8B_INSTRUCT_REFERENCE: &str = "meta-llama/Meta-Llama-3.1-8B-Instruct-Reference";
82pub const LLAMA_3_1_70B_INSTRUCT_REFERENCE: &str =
83    "meta-llama/Meta-Llama-3.1-70B-Instruct-Reference";
84pub const WIZARDLM_2_8X22B: &str = "microsoft/WizardLM-2-8x22B";
85pub const MISTRAL_7B_INSTRUCT_V0_1: &str = "mistralai/Mistral-7B-Instruct-v0.1";
86pub const MISTRAL_7B_INSTRUCT_V0_2: &str = "mistralai/Mistral-7B-Instruct-v0.2";
87pub const MISTRAL_7B_INSTRUCT_V0_3: &str = "mistralai/Mistral-7B-Instruct-v0.3";
88pub const MIXTRAL_8X7B_INSTRUCT_V0_1: &str = "mistralai/Mixtral-8x7B-Instruct-v0.1";
89pub const MIXTRAL_8X22B_INSTRUCT_V0_1: &str = "mistralai/Mixtral-8x22B-Instruct-v0.1";
90pub const NOUS_HERMES_2_MIXTRAL_8X7B_DPO: &str = "NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO";
91pub const NOUS_HERMES_LLAMA2_70B: &str = "NousResearch/Nous-Hermes-Llama2-70b";
92pub const NOUS_HERMES_2_MIXTRAL_8X7B_SFT: &str = "NousResearch/Nous-Hermes-2-Mixtral-8x7B-SFT";
93pub const NOUS_HERMES_LLAMA2_13B: &str = "NousResearch/Nous-Hermes-Llama2-13b";
94pub const NOUS_HERMES_2_MISTRAL_DPO: &str = "NousResearch/Nous-Hermes-2-Mistral-7B-DPO";
95pub const NOUS_HERMES_LLAMA2_7B: &str = "NousResearch/Nous-Hermes-llama-2-7b";
96pub const NOUS_CAPYBARA_V1_9: &str = "NousResearch/Nous-Capybara-7B-V1p9";
97pub const HERMES_2_THETA_LLAMA_3_70B: &str = "NousResearch/Hermes-2-Theta-Llama-3-70B";
98pub const OPENCHAT_3_5: &str = "openchat/openchat-3.5-1210";
99pub const OPENORCA_MISTRAL_7B_8K: &str = "Open-Orca/Mistral-7B-OpenOrca";
100pub const QWEN_2_72B_INSTRUCT: &str = "Qwen/Qwen2-72B-Instruct";
101pub const QWEN2_5_72B_INSTRUCT_TURBO: &str = "Qwen/Qwen2.5-72B-Instruct-Turbo";
102pub const QWEN2_5_7B_INSTRUCT_TURBO: &str = "Qwen/Qwen2.5-7B-Instruct-Turbo";
103pub const QWEN1_5_110B_CHAT: &str = "Qwen/Qwen1.5-110B-Chat";
104pub const QWEN1_5_72B_CHAT: &str = "Qwen/Qwen1.5-72B-Chat";
105pub const QWEN_2_1_5B_INSTRUCT: &str = "Qwen/Qwen2-1.5B-Instruct";
106pub const QWEN_2_7B_INSTRUCT: &str = "Qwen/Qwen2-7B-Instruct";
107pub const QWEN1_5_14B_CHAT: &str = "Qwen/Qwen1.5-14B-Chat";
108pub const QWEN1_5_1_8B_CHAT: &str = "Qwen/Qwen1.5-1.8B-Chat";
109pub const QWEN1_5_32B_CHAT: &str = "Qwen/Qwen1.5-32B-Chat";
110pub const QWEN1_5_7B_CHAT: &str = "Qwen/Qwen1.5-7B-Chat";
111pub const QWEN1_5_0_5B_CHAT: &str = "Qwen/Qwen1.5-0.5B-Chat";
112pub const QWEN1_5_4B_CHAT: &str = "Qwen/Qwen1.5-4B-Chat";
113pub const SNORKEL_MISTRAL_PAIRRM_DPO: &str = "snorkelai/Snorkel-Mistral-PairRM-DPO";
114pub const SNOWFLAKE_ARCTIC_INSTRUCT: &str = "Snowflake/snowflake-arctic-instruct";
115pub const ALPACA_7B: &str = "togethercomputer/alpaca-7b";
116pub const OPENHERMES_2_MISTRAL_7B: &str = "teknium/OpenHermes-2-Mistral-7B";
117pub const OPENHERMES_2_5_MISTRAL_7B: &str = "teknium/OpenHermes-2p5-Mistral-7B";
118pub const GUANACO_65B: &str = "togethercomputer/guanaco-65b";
119pub const GUANACO_13B: &str = "togethercomputer/guanaco-13b";
120pub const GUANACO_33B: &str = "togethercomputer/guanaco-33b";
121pub const GUANACO_7B: &str = "togethercomputer/guanaco-7b";
122pub const REMM_SLERP_L2_13B: &str = "Undi95/ReMM-SLERP-L2-13B";
123pub const TOPPY_M_7B: &str = "Undi95/Toppy-M-7B";
124pub const SOLAR_10_7B_INSTRUCT_V1: &str = "upstage/SOLAR-10.7B-Instruct-v1.0";
125pub const SOLAR_10_7B_INSTRUCT_V1_INT4: &str = "togethercomputer/SOLAR-10.7B-Instruct-v1.0-int4";
126pub const WIZARDLM_13B_V1_2: &str = "WizardLM/WizardLM-13B-V1.2";
127
128// =================================================================
129// Rig Implementation Types
130// =================================================================
131
132#[derive(Debug, Serialize, Deserialize)]
133pub(super) struct TogetherAICompletionRequest {
134    model: String,
135    pub messages: Vec<openai::Message>,
136    #[serde(flatten, skip_serializing_if = "Option::is_none")]
137    temperature: Option<f64>,
138    #[serde(skip_serializing_if = "Vec::is_empty")]
139    tools: Vec<crate::providers::openai::completion::ToolDefinition>,
140    #[serde(flatten, skip_serializing_if = "Option::is_none")]
141    tool_choice: Option<ToolChoice>,
142    #[serde(flatten, skip_serializing_if = "Option::is_none")]
143    pub additional_params: Option<serde_json::Value>,
144}
145
146impl TryFrom<(&str, CompletionRequest)> for TogetherAICompletionRequest {
147    type Error = CompletionError;
148
149    fn try_from((model, req): (&str, CompletionRequest)) -> Result<Self, Self::Error> {
150        let mut full_history: Vec<openai::Message> = match &req.preamble {
151            Some(preamble) => vec![openai::Message::system(preamble)],
152            None => vec![],
153        };
154        if let Some(docs) = req.normalized_documents() {
155            let docs: Vec<openai::Message> = docs.try_into()?;
156            full_history.extend(docs);
157        }
158
159        let chat_history: Vec<openai::Message> = req
160            .chat_history
161            .into_iter()
162            .map(|message| message.try_into())
163            .collect::<Result<Vec<Vec<openai::Message>>, _>>()?
164            .into_iter()
165            .flatten()
166            .collect();
167
168        full_history.extend(chat_history);
169
170        let tool_choice = req
171            .tool_choice
172            .clone()
173            .map(ToolChoice::try_from)
174            .transpose()?;
175
176        Ok(Self {
177            model: model.to_string(),
178            messages: full_history,
179            temperature: req.temperature,
180            tools: req
181                .tools
182                .clone()
183                .into_iter()
184                .map(crate::providers::openai::completion::ToolDefinition::from)
185                .collect::<Vec<_>>(),
186            tool_choice,
187            additional_params: req.additional_params,
188        })
189    }
190}
191
192#[derive(Clone)]
193pub struct CompletionModel<T = reqwest::Client> {
194    pub(crate) client: Client<T>,
195    pub model: String,
196}
197
198impl<T> CompletionModel<T> {
199    pub fn new(client: Client<T>, model: impl Into<String>) -> Self {
200        Self {
201            client,
202            model: model.into(),
203        }
204    }
205}
206
207impl<T> completion::CompletionModel for CompletionModel<T>
208where
209    T: HttpClientExt + Clone + Default + std::fmt::Debug + Send + 'static,
210{
211    type Response = openai::CompletionResponse;
212    type StreamingResponse = openai::StreamingCompletionResponse;
213
214    type Client = Client<T>;
215
216    fn make(client: &Self::Client, model: impl Into<String>) -> Self {
217        Self::new(client.clone(), model)
218    }
219
220    #[cfg_attr(feature = "worker", worker::send)]
221    async fn completion(
222        &self,
223        completion_request: completion::CompletionRequest,
224    ) -> Result<completion::CompletionResponse<openai::CompletionResponse>, CompletionError> {
225        let preamble = completion_request.preamble.clone();
226        let request = TogetherAICompletionRequest::try_from((
227            self.model.to_string().as_ref(),
228            completion_request,
229        ))?;
230        let messages_as_json_string = serde_json::to_string(&request.messages)?;
231
232        let span = if tracing::Span::current().is_disabled() {
233            info_span!(
234                target: "rig::completions",
235                "chat",
236                gen_ai.operation.name = "chat",
237                gen_ai.provider.name = "together",
238                gen_ai.request.model = self.model.to_string(),
239                gen_ai.system_instructions = preamble,
240                gen_ai.response.id = tracing::field::Empty,
241                gen_ai.response.model = tracing::field::Empty,
242                gen_ai.usage.output_tokens = tracing::field::Empty,
243                gen_ai.usage.input_tokens = tracing::field::Empty,
244                gen_ai.input.messages = &messages_as_json_string,
245                gen_ai.output.messages = tracing::field::Empty,
246            )
247        } else {
248            tracing::Span::current()
249        };
250
251        tracing::debug!(target: "rig::completions", "TogetherAI completion request: {messages_as_json_string}");
252
253        let body = serde_json::to_vec(&request)?;
254
255        let req = self
256            .client
257            .post("/v1/chat/completions")?
258            .body(body)
259            .map_err(|x| CompletionError::HttpError(x.into()))?;
260
261        async move {
262            let response = self.client.send::<_, Bytes>(req).await?;
263            let status = response.status();
264            let response_body = response.into_body().into_future().await?.to_vec();
265
266            if status.is_success() {
267                match serde_json::from_slice::<ApiResponse<openai::CompletionResponse>>(
268                    &response_body,
269                )? {
270                    ApiResponse::Ok(response) => {
271                        let span = tracing::Span::current();
272                        span.record(
273                            "gen_ai.output.messages",
274                            serde_json::to_string(&response.choices).unwrap(),
275                        );
276                        span.record("gen_ai.response.id", &response.id);
277                        span.record("gen_ai.response.model_name", &response.model);
278                        if let Some(ref usage) = response.usage {
279                            span.record("gen_ai.usage.input_tokens", usage.prompt_tokens);
280                            span.record(
281                                "gen_ai.usage.output_tokens",
282                                usage.total_tokens - usage.prompt_tokens,
283                            );
284                        }
285                        tracing::trace!(
286                            target: "rig::completions",
287                            "TogetherAI completion response: {}",
288                            serde_json::to_string_pretty(&response)?
289                        );
290                        response.try_into()
291                    }
292                    ApiResponse::Error(err) => Err(CompletionError::ProviderError(err.error)),
293                }
294            } else {
295                Err(CompletionError::ProviderError(
296                    String::from_utf8_lossy(&response_body).to_string(),
297                ))
298            }
299        }
300        .instrument(span)
301        .await
302    }
303
304    #[cfg_attr(feature = "worker", worker::send)]
305    async fn stream(
306        &self,
307        request: CompletionRequest,
308    ) -> Result<StreamingCompletionResponse<Self::StreamingResponse>, CompletionError> {
309        CompletionModel::stream(self, request).await
310    }
311}
312
313#[derive(Debug, Serialize, Deserialize)]
314#[serde(untagged, rename_all = "snake_case")]
315pub enum ToolChoice {
316    None,
317    Auto,
318    Function(Vec<ToolChoiceFunctionKind>),
319}
320
321impl TryFrom<crate::message::ToolChoice> for ToolChoice {
322    type Error = CompletionError;
323
324    fn try_from(value: crate::message::ToolChoice) -> Result<Self, Self::Error> {
325        let res = match value {
326            crate::message::ToolChoice::None => Self::None,
327            crate::message::ToolChoice::Auto => Self::Auto,
328            crate::message::ToolChoice::Specific { function_names } => {
329                let vec: Vec<ToolChoiceFunctionKind> = function_names
330                    .into_iter()
331                    .map(|name| ToolChoiceFunctionKind::Function { name })
332                    .collect();
333
334                Self::Function(vec)
335            }
336            choice => {
337                return Err(CompletionError::ProviderError(format!(
338                    "Unsupported tool choice type: {choice:?}"
339                )));
340            }
341        };
342
343        Ok(res)
344    }
345}
346
347#[derive(Debug, Serialize, Deserialize)]
348#[serde(tag = "type", content = "function")]
349pub enum ToolChoiceFunctionKind {
350    Function { name: String },
351}
rig/providers/together/completion.rs

rig/providers/together/
completion.rs
