openai_interface/moderations/
mod.rs

1//! Given text and/or image inputs, classifies if those inputs are potentially
2//! harmful.
3//!
4//! > ![warn] This module is untested!
5//! > OpenAI-compatible providers (DeepSeek, Qwen) are not known to implement
6//! > this endpoint, and no OpenAI API key was available for testing. If you
7//! > encounter any issues, please report them on the repository.
8
9pub mod request {
10    use serde::{Deserialize, Serialize};
11    use url::Url;
12
13    use crate::{
14        errors::OapiError,
15        rest::post::{Post, PostNoStream},
16    };
17
18    /// Classifies if text and/or image inputs are potentially harmful.
19    #[derive(Debug, Serialize, Deserialize, Default, Clone)]
20    pub struct ModerationRequest {
21        /// Input (or inputs) to classify. Can be a single string, an array of
22        /// strings, or an array of multi-modal input objects.
23        pub input: ModerationInput,
24        /// The content moderation model you would like to use, e.g.
25        /// `omni-moderation-latest`. Defaults to `omni-moderation-latest`
26        /// when not set.
27        #[serde(skip_serializing_if = "Option::is_none")]
28        pub model: Option<String>,
29        /// Add additional JSON properties to the request.
30        #[serde(flatten, default, skip_serializing_if = "Option::is_none")]
31        pub extra_body_map: Option<serde_json::Map<String, serde_json::Value>>,
32    }
33
34    /// The input (or inputs) to classify.
35    #[derive(Debug, Serialize, Deserialize, Clone)]
36    #[serde(untagged)]
37    pub enum ModerationInput {
38        /// A single string to classify.
39        String(String),
40        /// An array of strings to classify.
41        StringArray(Vec<String>),
42        /// An array of multi-modal input objects to classify.
43        MultiModal(Vec<ModerationInputContent>),
44    }
45
46    /// A multi-modal input object for moderation.
47    #[derive(Debug, Serialize, Deserialize, Clone)]
48    #[serde(tag = "type", rename_all = "snake_case")]
49    pub enum ModerationInputContent {
50        /// An object describing text to classify.
51        Text {
52            /// A string of text to classify.
53            text: String,
54        },
55        /// An object describing an image to classify.
56        ImageUrl {
57            /// Contains either an image URL or a data URL for a base64
58            /// encoded image.
59            image_url: ModerationImageUrl,
60        },
61    }
62
63    /// Contains either an image URL or a data URL for a base64 encoded
64    /// image.
65    #[derive(Debug, Serialize, Deserialize, Clone)]
66    pub struct ModerationImageUrl {
67        /// Either a URL of the image or the base64 encoded image data.
68        pub url: String,
69    }
70
71    impl Default for ModerationInput {
72        fn default() -> Self {
73            Self::String(String::new())
74        }
75    }
76
77    impl ModerationRequest {
78        pub fn is_streaming(&self) -> bool {
79            false
80        }
81    }
82
83    impl Post for ModerationRequest {
84        fn is_streaming(&self) -> bool {
85            false
86        }
87
88        /// Builds the URL for the request.
89        ///
90        /// `base_url` should be like <https://api.openai.com/v1>
91        fn build_url(&self, base_url: &str) -> Result<String, OapiError> {
92            let mut url =
93                Url::parse(base_url.trim_end_matches('/')).map_err(OapiError::UrlError)?;
94            url.path_segments_mut()
95                .map_err(|_| OapiError::UrlCannotBeBase(base_url.to_string()))?
96                .push("moderations");
97
98            Ok(url.to_string())
99        }
100    }
101
102    impl PostNoStream for ModerationRequest {
103        type Response = super::response::ModerationCreateResponse;
104    }
105}
106
107pub mod response {
108    use serde::{Deserialize, Serialize};
109
110    /// Represents if a given text input is potentially harmful.
111    #[derive(Debug, Deserialize, Serialize, Clone)]
112    pub struct ModerationCreateResponse {
113        /// The unique identifier for the moderation request.
114        pub id: String,
115        /// The model used to generate the moderation results.
116        pub model: String,
117        /// A list of moderation objects.
118        pub results: Vec<Moderation>,
119    }
120
121    /// A moderation object.
122    #[derive(Debug, Deserialize, Serialize, Clone)]
123    pub struct Moderation {
124        /// A list of the categories, and whether they are flagged or not.
125        pub categories: Categories,
126        /// A list of the categories along with the input type(s) that the
127        /// score applies to.
128        #[serde(alias = "category_applied_input_types")]
129        pub category_applied_input_types: Option<CategoryAppliedInputTypes>,
130        /// A list of the categories along with their scores as predicted by
131        /// the model.
132        #[serde(alias = "category_scores")]
133        pub category_scores: CategoryScores,
134        /// Whether any of the categories are flagged.
135        pub flagged: bool,
136    }
137
138    /// A list of the categories, and whether they are flagged or not.
139    ///
140    /// Fields are optional so that responses of different moderation model
141    /// generations (which add or remove categories over time) can be
142    /// deserialized.
143    #[derive(Debug, Deserialize, Serialize, Clone, Default)]
144    pub struct Categories {
145        pub harassment: Option<bool>,
146        #[serde(rename = "harassment/threatening", alias = "harassment_threatening")]
147        pub harassment_threatening: Option<bool>,
148        pub hate: Option<bool>,
149        #[serde(rename = "hate/threatening", alias = "hate_threatening")]
150        pub hate_threatening: Option<bool>,
151        pub illicit: Option<bool>,
152        #[serde(rename = "illicit/violent", alias = "illicit_violent")]
153        pub illicit_violent: Option<bool>,
154        #[serde(rename = "self-harm", alias = "self_harm")]
155        pub self_harm: Option<bool>,
156        #[serde(rename = "self-harm/instructions", alias = "self_harm_instructions")]
157        pub self_harm_instructions: Option<bool>,
158        #[serde(rename = "self-harm/intent", alias = "self_harm_intent")]
159        pub self_harm_intent: Option<bool>,
160        pub sexual: Option<bool>,
161        #[serde(rename = "sexual/minors", alias = "sexual_minors")]
162        pub sexual_minors: Option<bool>,
163        pub violence: Option<bool>,
164        #[serde(rename = "violence/graphic", alias = "violence_graphic")]
165        pub violence_graphic: Option<bool>,
166    }
167
168    /// A list of the categories along with the input type(s) that the score
169    /// applies to, e.g. `["text"]` or `["text", "image"]`.
170    #[derive(Debug, Deserialize, Serialize, Clone, Default)]
171    pub struct CategoryAppliedInputTypes {
172        pub harassment: Option<Vec<String>>,
173        #[serde(rename = "harassment/threatening", alias = "harassment_threatening")]
174        pub harassment_threatening: Option<Vec<String>>,
175        pub hate: Option<Vec<String>>,
176        #[serde(rename = "hate/threatening", alias = "hate_threatening")]
177        pub hate_threatening: Option<Vec<String>>,
178        pub illicit: Option<Vec<String>>,
179        #[serde(rename = "illicit/violent", alias = "illicit_violent")]
180        pub illicit_violent: Option<Vec<String>>,
181        #[serde(rename = "self-harm", alias = "self_harm")]
182        pub self_harm: Option<Vec<String>>,
183        #[serde(rename = "self-harm/instructions", alias = "self_harm_instructions")]
184        pub self_harm_instructions: Option<Vec<String>>,
185        #[serde(rename = "self-harm/intent", alias = "self_harm_intent")]
186        pub self_harm_intent: Option<Vec<String>>,
187        pub sexual: Option<Vec<String>>,
188        #[serde(rename = "sexual/minors", alias = "sexual_minors")]
189        pub sexual_minors: Option<Vec<String>>,
190        pub violence: Option<Vec<String>>,
191        #[serde(rename = "violence/graphic", alias = "violence_graphic")]
192        pub violence_graphic: Option<Vec<String>>,
193    }
194
195    /// A list of the categories along with their scores as predicted by the
196    /// model.
197    #[derive(Debug, Deserialize, Serialize, Clone, Default)]
198    pub struct CategoryScores {
199        pub harassment: Option<f32>,
200        #[serde(rename = "harassment/threatening", alias = "harassment_threatening")]
201        pub harassment_threatening: Option<f32>,
202        pub hate: Option<f32>,
203        #[serde(rename = "hate/threatening", alias = "hate_threatening")]
204        pub hate_threatening: Option<f32>,
205        pub illicit: Option<f32>,
206        #[serde(rename = "illicit/violent", alias = "illicit_violent")]
207        pub illicit_violent: Option<f32>,
208        #[serde(rename = "self-harm", alias = "self_harm")]
209        pub self_harm: Option<f32>,
210        #[serde(rename = "self-harm/instructions", alias = "self_harm_instructions")]
211        pub self_harm_instructions: Option<f32>,
212        #[serde(rename = "self-harm/intent", alias = "self_harm_intent")]
213        pub self_harm_intent: Option<f32>,
214        pub sexual: Option<f32>,
215        #[serde(rename = "sexual/minors", alias = "sexual_minors")]
216        pub sexual_minors: Option<f32>,
217        pub violence: Option<f32>,
218        #[serde(rename = "violence/graphic", alias = "violence_graphic")]
219        pub violence_graphic: Option<f32>,
220    }
221
222    crate::impl_from_str!(ModerationCreateResponse);
223}
224
225#[cfg(test)]
226mod tests {
227    use super::request::{
228        ModerationImageUrl, ModerationInput, ModerationInputContent, ModerationRequest,
229    };
230    use crate::rest::post::Post;
231
232    /// Serializes a simple text moderation request.
233    #[test]
234    fn request_serialization() {
235        let request = ModerationRequest {
236            input: ModerationInput::String("I want to harm someone.".to_string()),
237            model: Some("omni-moderation-latest".to_string()),
238            ..Default::default()
239        };
240
241        let json = serde_json::to_string(&request).unwrap();
242        assert!(
243            json.contains(r#""input":"I want to harm someone.""#),
244            "json: {json}"
245        );
246        assert!(
247            json.contains(r#""model":"omni-moderation-latest""#),
248            "json: {json}"
249        );
250    }
251
252    /// Serializes a multi-modal moderation input array with the official
253    /// `text` / `image_url` shapes.
254    #[test]
255    fn multimodal_input_serialization() {
256        let request = ModerationRequest {
257            input: ModerationInput::MultiModal(vec![
258                ModerationInputContent::Text {
259                    text: "hello".to_string(),
260                },
261                ModerationInputContent::ImageUrl {
262                    image_url: ModerationImageUrl {
263                        url: "https://example.com/a.png".to_string(),
264                    },
265                },
266            ]),
267            ..Default::default()
268        };
269
270        let json = serde_json::to_string(&request).unwrap();
271        assert!(
272            json.contains(r#"{"type":"text","text":"hello"}"#),
273            "json: {json}"
274        );
275        assert!(
276            json.contains(
277                r#"{"type":"image_url","image_url":{"url":"https://example.com/a.png"}}"#
278            ),
279            "json: {json}"
280        );
281    }
282
283    #[test]
284    fn test_build_url() {
285        let request = ModerationRequest::default();
286        let url = request.build_url("https://api.openai.com/v1/").unwrap();
287        assert_eq!(url, "https://api.openai.com/v1/moderations");
288    }
289
290    /// Deserializes a moderation response.
291    ///
292    /// No accessible provider implements this endpoint, so this fixture is
293    /// NOT captured from a live response. The structure (category names,
294    /// slash-separated keys, aliases, `flagged`) follows the schema of
295    /// openai-python `types/moderation.py`; the values are constructed for
296    /// the test. Fields absent from the fixture exercise the optional-field
297    /// tolerance of the types.
298    #[test]
299    fn parse_response() {
300        let content = r#"{
301            "id": "modr-0a1b2c3d4e5f6a7b8c9d0e1f",
302            "model": "omni-moderation-latest",
303            "results": [
304                {
305                    "flagged": true,
306                    "categories": {
307                        "harassment": false,
308                        "harassment/threatening": false,
309                        "sexual": false,
310                        "self-harm/instructions": true,
311                        "violence": true
312                    },
313                    "category_scores": {
314                        "harassment": 0.01,
315                        "harassment/threatening": 0.02,
316                        "sexual": 0.0001,
317                        "self-harm/instructions": 0.87,
318                        "violence": 0.93
319                    },
320                    "category_applied_input_types": {
321                        "sexual": ["text"],
322                        "violence": ["text", "image"]
323                    }
324                }
325            ]
326        }"#;
327
328        let response: super::response::ModerationCreateResponse = content.parse().unwrap();
329        assert_eq!(response.results.len(), 1);
330
331        let result = &response.results[0];
332        assert!(result.flagged);
333        assert_eq!(result.categories.violence, Some(true));
334        assert_eq!(result.categories.self_harm_instructions, Some(true));
335        assert_eq!(result.categories.harassment, Some(false));
336        assert_eq!(result.category_scores.violence, Some(0.93));
337        assert_eq!(result.category_scores.self_harm_instructions, Some(0.87));
338        let applied = result.category_applied_input_types.as_ref().unwrap();
339        assert_eq!(
340            applied.violence,
341            Some(vec!["text".to_string(), "image".to_string()])
342        );
343    }
344}
openai_interface/moderations/mod.rs

openai_interface/moderations/
mod.rs
