llamaedge/
params.rs

1//! Parameters for the chat completion API.
2
3#[cfg(feature = "audio")]
4use endpoints::audio::transcription::TimestampGranularity;
5use endpoints::chat::{ChatResponseFormat, Tool, ToolChoice};
6#[cfg(feature = "image")]
7use endpoints::{
8    files::FileObject,
9    images::{SamplingMethod, Scheduler},
10};
11#[cfg(feature = "image")]
12use std::path::PathBuf;
13
14#[cfg(feature = "image")]
15pub type ImageResponseFormat = endpoints::images::ResponseFormat;
16
17/// Parameters for the chat completion API.
18#[derive(Debug, Clone)]
19pub struct ChatParams {
20    /// The model to use for generating completions.
21    pub model: Option<String>,
22    /// Adjust the randomness of the generated text. Between 0.0 and 2.0. Higher values like 0.8 will make the output more random, while lower values like 0.2 will make it more focused and deterministic.
23    ///
24    /// We generally recommend altering this or top_p but not both.
25    /// Defaults to 1.0.
26    pub temperature: Option<f64>,
27    /// Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P. The value should be between 0.0 and 1.0.
28    ///
29    /// Top-p sampling, also known as nucleus sampling, is another text generation method that selects the next token from a subset of tokens that together have a cumulative probability of at least p. This method provides a balance between diversity and quality by considering both the probabilities of tokens and the number of tokens to sample from. A higher value for top_p (e.g., 0.95) will lead to more diverse text, while a lower value (e.g., 0.5) will generate more focused and conservative text.
30    ///
31    /// We generally recommend altering this or temperature but not both.
32    /// Defaults to 1.0.
33    pub top_p: Option<f64>,
34    /// How many chat completion choices to generate for each input message.
35    /// Defaults to 1.
36    pub n_choice: Option<u64>,
37    /// A list of tokens at which to stop generation. If None, no stop tokens are used. Up to 4 sequences where the API will stop generating further tokens.
38    /// Defaults to None
39    pub stop: Option<Vec<String>>,
40    /// **Deprecated** Use `max_completion_tokens` instead.
41    ///
42    /// The maximum number of tokens to generate. The value should be no less than 1.
43    /// Defaults to 1024.
44    pub max_tokens: Option<u64>,
45    /// An upper bound for the number of tokens that can be generated for a completion. Defaults to -1, which means no upper bound.
46    pub max_completion_tokens: Option<i32>,
47    /// Number between -2.0 and 2.0. Positive values penalize new tokens based on whether they appear in the text so far, increasing the model's likelihood to talk about new topics.
48    /// Defaults to 0.0.
49    pub presence_penalty: Option<f64>,
50    /// Number between -2.0 and 2.0. Positive values penalize new tokens based on their existing frequency in the text so far, decreasing the model's likelihood to repeat the same line verbatim.
51    /// Defaults to 0.0.
52    pub frequency_penalty: Option<f64>,
53    /// A unique identifier representing your end-user.
54    pub user: Option<String>,
55    /// Format that the model must output
56    pub response_format: Option<ChatResponseFormat>,
57    /// A list of tools the model may call.
58    ///
59    /// Currently, only functions are supported as a tool. Use this to provide a list of functions the model may generate JSON inputs for.
60    pub tools: Option<Vec<Tool>>,
61    /// Controls which (if any) function is called by the model.
62    pub tool_choice: Option<ToolChoice>,
63}
64impl Default for ChatParams {
65    fn default() -> Self {
66        Self {
67            model: None,
68            temperature: Some(1.0),
69            top_p: Some(1.0),
70            n_choice: Some(1),
71            stop: None,
72            max_tokens: Some(1024),
73            max_completion_tokens: Some(-1),
74            presence_penalty: Some(0.0),
75            frequency_penalty: Some(0.0),
76            user: None,
77            response_format: None,
78            tools: None,
79            tool_choice: None,
80        }
81    }
82}
83
84/// Parameters for the RAG chat completion API.
85#[cfg(feature = "rag")]
86#[derive(Debug, Clone)]
87pub struct RagChatParams {
88    /// The model to use for generating completions.
89    pub model: Option<String>,
90    /// Adjust the randomness of the generated text. Between 0.0 and 2.0. Higher values like 0.8 will make the output more random, while lower values like 0.2 will make it more focused and deterministic.
91    ///
92    /// We generally recommend altering this or top_p but not both.
93    /// Defaults to 0.8.
94    pub temperature: f64,
95    /// Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P. The value should be between 0.0 and 1.0.
96    ///
97    /// Top-p sampling, also known as nucleus sampling, is another text generation method that selects the next token from a subset of tokens that together have a cumulative probability of at least p. This method provides a balance between diversity and quality by considering both the probabilities of tokens and the number of tokens to sample from. A higher value for top_p (e.g., 0.95) will lead to more diverse text, while a lower value (e.g., 0.5) will generate more focused and conservative text.
98    ///
99    /// We generally recommend altering this or temperature but not both.
100    /// Defaults to 0.9. To disable top-p sampling, set it to 1.0.
101    pub top_p: f64,
102    /// How many chat completion choices to generate for each input message.
103    /// Defaults to 1.
104    pub n_choice: u64,
105    /// A list of tokens at which to stop generation. If None, no stop tokens are used. Up to 4 sequences where the API will stop generating further tokens.
106    /// Defaults to None
107    pub stop: Option<Vec<String>>,
108    /// **Deprecated** Use `max_completion_tokens` instead.
109    ///
110    /// The maximum number of tokens to generate. The value should be no less than 1.
111    /// Defaults to 1024.
112    pub max_tokens: u64,
113    /// An upper bound for the number of tokens that can be generated for a completion. Defaults to -1, which means no upper bound.
114    pub max_completion_tokens: i32,
115    /// Number between -2.0 and 2.0. Positive values penalize new tokens based on whether they appear in the text so far, increasing the model's likelihood to talk about new topics.
116    /// Defaults to 0.0.
117    pub presence_penalty: f64,
118    /// Number between -2.0 and 2.0. Positive values penalize new tokens based on their existing frequency in the text so far, decreasing the model's likelihood to repeat the same line verbatim.
119    /// Defaults to 0.0.
120    pub frequency_penalty: f64,
121    /// A unique identifier representing your end-user.
122    pub user: Option<String>,
123    /// Format that the model must output
124    pub response_format: Option<ChatResponseFormat>,
125    /// A list of tools the model may call.
126    ///
127    /// Currently, only functions are supported as a tool. Use this to provide a list of functions the model may generate JSON inputs for.
128    pub tools: Option<Vec<Tool>>,
129    /// Controls which (if any) function is called by the model.
130    pub tool_choice: Option<ToolChoice>,
131    /// Number of user messages to use for context retrieval. Defaults to 1.
132    pub context_window: u64,
133    /// The configuration for the VectorDB server.
134    pub vdb_config: Option<RagVdbConfig>,
135}
136#[cfg(feature = "rag")]
137impl Default for RagChatParams {
138    fn default() -> Self {
139        Self {
140            model: None,
141            temperature: 0.8,
142            top_p: 0.9,
143            n_choice: 1,
144            stop: None,
145            max_tokens: 1024,
146            max_completion_tokens: -1,
147            presence_penalty: 0.0,
148            frequency_penalty: 0.0,
149            user: None,
150            response_format: None,
151            tools: None,
152            tool_choice: None,
153            context_window: 1,
154            vdb_config: None,
155        }
156    }
157}
158
159/// The configuration for the VectorDB server.
160#[cfg(feature = "rag")]
161#[derive(Debug, Clone)]
162pub struct RagVdbConfig {
163    /// The URL of the VectorDB server.
164    pub server_url: String,
165    /// The names of the collections in VectorDB.
166    pub collection_name: Vec<String>,
167    /// Max number of retrieved results. The number of the values must be the same as the number of `collection_name`.
168    pub limit: Vec<u64>,
169    /// The score threshold for the retrieved results. The number of the values must be the same as the number of `collection_name`.
170    pub score_threshold: Vec<f32>,
171    /// The API key for the VectorDB server.
172    pub api_key: Option<String>,
173}
174
175/// Parameters for the transcription API.
176#[cfg(feature = "audio")]
177#[derive(Debug, Clone)]
178pub struct TranscriptionParams {
179    /// ID of the model to use.
180    pub model: Option<String>,
181    /// An optional text to guide the model's style or continue a previous audio segment. The prompt should match the audio language.
182    pub prompt: Option<String>,
183    /// The format of the transcript output, in one of these options: `json`, `text`, `srt`, `verbose_json`, or `vtt`.
184    pub response_format: String,
185    /// The sampling temperature, between 0 and 1. Higher values like 0.8 will make the output more random, while lower values like 0.2 will make it more focused and deterministic. If set to 0, the model will use [log probability](https://en.wikipedia.org/wiki/Log_probability) to automatically increase the temperature until certain thresholds are hit. Defaults to 0.0.
186    pub temperature: f64,
187    /// The timestamp granularities to populate for this transcription.
188    /// `response_format` must be set `verbose_json` to use timestamp granularities. Either or both of these options are supported: `word`, or `segment`.
189    pub timestamp_granularities: Option<Vec<TimestampGranularity>>,
190
191    /// Automatically detect the spoken language in the provided audio input. Defaults to false.
192    pub detect_language: bool,
193    /// Time offset in milliseconds. Defaults to 0.
194    pub offset_time: u64,
195    /// Length of audio (in seconds) to be processed starting from the point defined by the `offset_time` field (or from the beginning by default). Defaults to 0.
196    pub duration: u64,
197    /// Maximum amount of text context (in tokens) that the model uses when processing long audio inputs incrementally. Defaults to -1.
198    pub max_context: i32,
199    /// Maximum number of tokens that the model can generate in a single transcription segment (or chunk). Defaults to 0.
200    pub max_len: u64,
201    /// Split audio chunks on word rather than on token. Defaults to false.
202    pub split_on_word: bool,
203    /// Use the new computation context. Defaults to false.
204    pub use_new_context: bool,
205}
206#[cfg(feature = "audio")]
207impl Default for TranscriptionParams {
208    fn default() -> Self {
209        Self {
210            model: None,
211            prompt: None,
212            response_format: "json".to_string(),
213            temperature: 0.0,
214            timestamp_granularities: Some(vec![TimestampGranularity::Segment]),
215            detect_language: false,
216            offset_time: 0,
217            duration: 0,
218            max_context: -1,
219            max_len: 0,
220            split_on_word: false,
221            use_new_context: false,
222        }
223    }
224}
225
226/// Parameters for the translation API.
227#[cfg(feature = "audio")]
228#[derive(Debug, Clone)]
229pub struct TranslationParams {
230    /// ID of the model to use.
231    pub model: Option<String>,
232    /// An optional text to guide the model's style or continue a previous audio segment. The prompt should be in English.
233    pub prompt: Option<String>,
234    /// The format of the transcript output, in one of these options: `json`, `text`, `srt`, `verbose_json`, or `vtt`. Defaults to `json`.
235    pub response_format: String,
236    /// The sampling temperature, between 0 and 1. Higher values like 0.8 will make the output more random, while lower values like 0.2 will make it more focused and deterministic. If set to 0, the model will use [log probability](https://en.wikipedia.org/wiki/Log_probability) to automatically increase the temperature until certain thresholds are hit. Defaults to 0.0.
237    pub temperature: f64,
238    /// automatically detect the spoken language in the provided audio input. Defaults to false.
239    pub detect_language: bool,
240    /// Time offset in milliseconds. Defaults to 0.
241    pub offset_time: u64,
242    /// Length of audio (in seconds) to be processed starting from the point defined by the `offset_time` field (or from the beginning by default). Defaults to 0.
243    pub duration: u64,
244    /// Maximum amount of text context (in tokens) that the model uses when processing long audio inputs incrementally. Defaults to -1.
245    pub max_context: i32,
246    /// Maximum number of tokens that the model can generate in a single transcription segment (or chunk). Defaults to 0.
247    pub max_len: u64,
248    /// Split audio chunks on word rather than on token. Defaults to false.
249    pub split_on_word: bool,
250    /// Use the new computation context. Defaults to false.
251    pub use_new_context: bool,
252}
253#[cfg(feature = "audio")]
254impl Default for TranslationParams {
255    fn default() -> Self {
256        Self {
257            model: None,
258            prompt: None,
259            response_format: "json".to_string(),
260            temperature: 0.0,
261            detect_language: false,
262            offset_time: 0,
263            duration: 0,
264            max_context: -1,
265            max_len: 0,
266            split_on_word: false,
267            use_new_context: false,
268        }
269    }
270}
271
272/// Parameters for the embeddings API.
273#[derive(Debug, Clone)]
274pub struct EmbeddingsParams {
275    /// ID of the model to use.
276    pub model: Option<String>,
277    /// The format to return the embeddings in. Can be either float or base64.
278    /// Defaults to float.
279    pub encoding_format: String,
280    /// A unique identifier representing your end-user.
281    pub user: Option<String>,
282    /// The URL of the VectorDB server.
283    #[cfg(feature = "rag")]
284    pub vdb_server_url: Option<String>,
285    /// The name of the collection in VectorDB.
286    #[cfg(feature = "rag")]
287    pub vdb_collection_name: Option<String>,
288    /// The API key for the VectorDB server.
289    #[cfg(feature = "rag")]
290    pub vdb_api_key: Option<String>,
291}
292impl Default for EmbeddingsParams {
293    fn default() -> Self {
294        Self {
295            model: None,
296            encoding_format: "float".to_string(),
297            user: None,
298            #[cfg(feature = "rag")]
299            vdb_server_url: None,
300            #[cfg(feature = "rag")]
301            vdb_collection_name: None,
302            #[cfg(feature = "rag")]
303            vdb_api_key: None,
304        }
305    }
306}
307
308/// Parameters for the image generation API.
309#[cfg(feature = "image")]
310#[derive(Debug, Clone)]
311pub struct ImageCreateParams {
312    /// Negative prompt for the image generation.
313    pub negative_prompt: Option<String>,
314    /// Name of the model to use for image generation.
315    pub model: String,
316    /// Number of images to generate. Defaults to `1`.
317    pub n: u64,
318    /// The format in which the generated images are returned. Must be one of `url` or `b64_json`. Defaults to `Url`.
319    pub response_format: ImageResponseFormat,
320    /// A unique identifier representing your end-user, which can help monitor and detect abuse.
321    pub user: Option<String>,
322    /// Unconditional guidance scale. Defaults to `7.0`.
323    pub cfg_scale: f32,
324    /// Sampling method. Defaults to "euler_a".
325    pub sample_method: SamplingMethod,
326    /// Number of sample steps. Defaults to `20`.
327    pub steps: usize,
328    /// Image height, in pixel space. Defaults to `512`.
329    pub height: usize,
330    /// Image width, in pixel space. Defaults to `512`.
331    pub width: usize,
332    /// Strength to apply Control Net. Defaults to `0.9`.
333    pub control_strength: f32,
334    /// The image to control the generation.
335    pub control_image: Option<FileObject>,
336    /// RNG seed. Negative value means to use random seed. Defaults to `42`.
337    pub seed: i32,
338    /// Strength for noising/unnoising. Defaults to `0.75`.
339    pub strength: f32,
340    /// Denoiser sigma scheduler. Possible values are `discrete`, `karras`, `exponential`, `ays`, `gits`. Defaults to `discrete`.
341    pub scheduler: Scheduler,
342    /// Apply canny preprocessor. Defaults to `false`.
343    pub apply_canny_preprocessor: bool,
344    /// Strength for keeping input identity. Defaults to `0.2`.
345    pub style_ratio: f32,
346}
347#[cfg(feature = "image")]
348impl Default for ImageCreateParams {
349    fn default() -> Self {
350        Self {
351            negative_prompt: None,
352            model: "".to_string(),
353            n: 1,
354            response_format: ImageResponseFormat::Url,
355            user: None,
356            cfg_scale: 7.0,
357            sample_method: SamplingMethod::EulerA,
358            steps: 20,
359            height: 512,
360            width: 512,
361            control_strength: 0.9,
362            control_image: None,
363            seed: 42,
364            strength: 0.75,
365            scheduler: Scheduler::Discrete,
366            apply_canny_preprocessor: false,
367            style_ratio: 0.2,
368        }
369    }
370}
371
372/// Parameters for the image edit API.
373#[cfg(feature = "image")]
374#[derive(Debug, Clone)]
375pub struct ImageEditParams {
376    /// Negative prompt for the image generation.
377    pub negative_prompt: Option<String>,
378    /// An additional image whose fully transparent areas (e.g. where alpha is zero) indicate where `image` should be edited. Must have the same dimensions as the input image.
379    pub mask: Option<PathBuf>,
380    /// The model to use for image generation.
381    pub model: String,
382    /// The number of images to generate. Defaults to `1`.
383    pub n: u64,
384    /// The format in which the generated images are returned. Must be one of `url` or `b64_json`. Defaults to `url`.
385    pub response_format: ImageResponseFormat,
386    /// A unique identifier representing your end-user, which can help monitor and detect abuse.
387    pub user: Option<String>,
388    /// Unconditional guidance scale. Defaults to `7.0`.
389    pub cfg_scale: f32,
390    /// Sampling method. Defaults to "euler_a".
391    pub sample_method: SamplingMethod,
392    /// Number of sample steps. Defaults to `20`.
393    pub steps: usize,
394    /// Image height, in pixel space. Defaults to `512`.
395    pub height: usize,
396    /// Image width, in pixel space. Defaults to `512`.
397    pub width: usize,
398    /// strength to apply Control Net. Defaults to `0.9`.
399    pub control_strength: f32,
400    /// The image to control the generation.
401    pub control_image: Option<PathBuf>,
402    /// RNG seed. Negative value means to use random seed. Defaults to `42`.
403    pub seed: i32,
404    /// Strength for noising/unnoising. Defaults to `0.75`.
405    pub strength: f32,
406    /// Denoiser sigma scheduler. Possible values are `discrete`, `karras`, `exponential`, `ays`, `gits`. Defaults to `discrete`.
407    pub scheduler: Scheduler,
408    /// Apply canny preprocessor. Defaults to `false`.
409    pub apply_canny_preprocessor: bool,
410    /// Strength for keeping input identity. Defaults to `0.2`.
411    pub style_ratio: f32,
412}
413#[cfg(feature = "image")]
414impl Default for ImageEditParams {
415    fn default() -> Self {
416        Self {
417            negative_prompt: None,
418            mask: None,
419            model: "".to_string(),
420            n: 1,
421            response_format: ImageResponseFormat::Url,
422            user: None,
423            cfg_scale: 7.0,
424            sample_method: SamplingMethod::EulerA,
425            steps: 20,
426            height: 512,
427            width: 512,
428            control_strength: 0.9,
429            control_image: None,
430            seed: 42,
431            strength: 0.75,
432            scheduler: Scheduler::Discrete,
433            apply_canny_preprocessor: false,
434            style_ratio: 0.2,
435        }
436    }
437}