1use std::io::{Cursor, Write};
2use std::sync::Arc;
3
4use anyhow::Result;
5use base64::write::EncoderWriter;
6use cloud_llm_client::{CompletionIntent, CompletionMode};
7use gpui::{
8 App, AppContext as _, DevicePixels, Image, ImageFormat, ObjectFit, SharedString, Size, Task,
9 point, px, size,
10};
11use image::codecs::png::PngEncoder;
12use serde::{Deserialize, Serialize};
13use util::ResultExt;
14
15use crate::role::Role;
16use crate::{LanguageModelToolUse, LanguageModelToolUseId};
17
18#[derive(Clone, PartialEq, Eq, Serialize, Deserialize, Hash)]
19pub struct LanguageModelImage {
20 /// A base64-encoded PNG image.
21 pub source: SharedString,
22 pub size: Size<DevicePixels>,
23}
24
25impl LanguageModelImage {
26 pub fn len(&self) -> usize {
27 self.source.len()
28 }
29
30 pub fn is_empty(&self) -> bool {
31 self.source.is_empty()
32 }
33
34 // Parse Self from a JSON object with case-insensitive field names
35 pub fn from_json(obj: &serde_json::Map<String, serde_json::Value>) -> Option<Self> {
36 let mut source = None;
37 let mut size_obj = None;
38
39 // Find source and size fields (case-insensitive)
40 for (k, v) in obj.iter() {
41 match k.to_lowercase().as_str() {
42 "source" => source = v.as_str(),
43 "size" => size_obj = v.as_object(),
44 _ => {}
45 }
46 }
47
48 let source = source?;
49 let size_obj = size_obj?;
50
51 let mut width = None;
52 let mut height = None;
53
54 // Find width and height in size object (case-insensitive)
55 for (k, v) in size_obj.iter() {
56 match k.to_lowercase().as_str() {
57 "width" => width = v.as_i64().map(|w| w as i32),
58 "height" => height = v.as_i64().map(|h| h as i32),
59 _ => {}
60 }
61 }
62
63 Some(Self {
64 size: size(DevicePixels(width?), DevicePixels(height?)),
65 source: SharedString::from(source.to_string()),
66 })
67 }
68}
69
70impl std::fmt::Debug for LanguageModelImage {
71 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
72 f.debug_struct("LanguageModelImage")
73 .field("source", &format!("<{} bytes>", self.source.len()))
74 .field("size", &self.size)
75 .finish()
76 }
77}
78
79/// Anthropic wants uploaded images to be smaller than this in both dimensions.
80const ANTHROPIC_SIZE_LIMT: f32 = 1568.;
81
82impl LanguageModelImage {
83 pub fn empty() -> Self {
84 Self {
85 source: "".into(),
86 size: size(DevicePixels(0), DevicePixels(0)),
87 }
88 }
89
90 pub fn from_image(data: Arc<Image>, cx: &mut App) -> Task<Option<Self>> {
91 cx.background_spawn(async move {
92 let image_bytes = Cursor::new(data.bytes());
93 let dynamic_image = match data.format() {
94 ImageFormat::Png => image::codecs::png::PngDecoder::new(image_bytes)
95 .and_then(image::DynamicImage::from_decoder),
96 ImageFormat::Jpeg => image::codecs::jpeg::JpegDecoder::new(image_bytes)
97 .and_then(image::DynamicImage::from_decoder),
98 ImageFormat::Webp => image::codecs::webp::WebPDecoder::new(image_bytes)
99 .and_then(image::DynamicImage::from_decoder),
100 ImageFormat::Gif => image::codecs::gif::GifDecoder::new(image_bytes)
101 .and_then(image::DynamicImage::from_decoder),
102 _ => return None,
103 }
104 .log_err()?;
105
106 let width = dynamic_image.width();
107 let height = dynamic_image.height();
108 let image_size = size(DevicePixels(width as i32), DevicePixels(height as i32));
109
110 let base64_image = {
111 if image_size.width.0 > ANTHROPIC_SIZE_LIMT as i32
112 || image_size.height.0 > ANTHROPIC_SIZE_LIMT as i32
113 {
114 let new_bounds = ObjectFit::ScaleDown.get_bounds(
115 gpui::Bounds {
116 origin: point(px(0.0), px(0.0)),
117 size: size(px(ANTHROPIC_SIZE_LIMT), px(ANTHROPIC_SIZE_LIMT)),
118 },
119 image_size,
120 );
121 let resized_image = dynamic_image.resize(
122 new_bounds.size.width.0 as u32,
123 new_bounds.size.height.0 as u32,
124 image::imageops::FilterType::Triangle,
125 );
126
127 encode_as_base64(data, resized_image)
128 } else {
129 encode_as_base64(data, dynamic_image)
130 }
131 }
132 .log_err()?;
133
134 // SAFETY: The base64 encoder should not produce non-UTF8.
135 let source = unsafe { String::from_utf8_unchecked(base64_image) };
136
137 Some(LanguageModelImage {
138 size: image_size,
139 source: source.into(),
140 })
141 })
142 }
143
144 pub fn estimate_tokens(&self) -> usize {
145 let width = self.size.width.0.unsigned_abs() as usize;
146 let height = self.size.height.0.unsigned_abs() as usize;
147
148 // From: https://docs.anthropic.com/en/docs/build-with-claude/vision#calculate-image-costs
149 // Note that are a lot of conditions on Anthropic's API, and OpenAI doesn't use this,
150 // so this method is more of a rough guess.
151 (width * height) / 750
152 }
153
154 pub fn to_base64_url(&self) -> String {
155 format!("data:image/png;base64,{}", self.source)
156 }
157}
158
159fn encode_as_base64(data: Arc<Image>, image: image::DynamicImage) -> Result<Vec<u8>> {
160 let mut base64_image = Vec::new();
161 {
162 let mut base64_encoder = EncoderWriter::new(
163 Cursor::new(&mut base64_image),
164 &base64::engine::general_purpose::STANDARD,
165 );
166 if data.format() == ImageFormat::Png {
167 base64_encoder.write_all(data.bytes())?;
168 } else {
169 let mut png = Vec::new();
170 image.write_with_encoder(PngEncoder::new(&mut png))?;
171
172 base64_encoder.write_all(png.as_slice())?;
173 }
174 }
175 Ok(base64_image)
176}
177
178#[derive(Debug, Clone, Serialize, Deserialize, Eq, PartialEq, Hash)]
179pub struct LanguageModelToolResult {
180 pub tool_use_id: LanguageModelToolUseId,
181 pub tool_name: Arc<str>,
182 pub is_error: bool,
183 pub content: LanguageModelToolResultContent,
184 pub output: Option<serde_json::Value>,
185}
186
187#[derive(Debug, Clone, Serialize, Eq, PartialEq, Hash)]
188pub enum LanguageModelToolResultContent {
189 Text(Arc<str>),
190 Image(LanguageModelImage),
191}
192
193impl<'de> Deserialize<'de> for LanguageModelToolResultContent {
194 fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
195 where
196 D: serde::Deserializer<'de>,
197 {
198 use serde::de::Error;
199
200 let value = serde_json::Value::deserialize(deserializer)?;
201
202 // Models can provide these responses in several styles. Try each in order.
203
204 // 1. Try as plain string
205 if let Ok(text) = serde_json::from_value::<String>(value.clone()) {
206 return Ok(Self::Text(Arc::from(text)));
207 }
208
209 // 2. Try as object
210 if let Some(obj) = value.as_object() {
211 // get a JSON field case-insensitively
212 fn get_field<'a>(
213 obj: &'a serde_json::Map<String, serde_json::Value>,
214 field: &str,
215 ) -> Option<&'a serde_json::Value> {
216 obj.iter()
217 .find(|(k, _)| k.to_lowercase() == field.to_lowercase())
218 .map(|(_, v)| v)
219 }
220
221 // Accept wrapped text format: { "type": "text", "text": "..." }
222 if let (Some(type_value), Some(text_value)) =
223 (get_field(obj, "type"), get_field(obj, "text"))
224 && let Some(type_str) = type_value.as_str()
225 && type_str.to_lowercase() == "text"
226 && let Some(text) = text_value.as_str()
227 {
228 return Ok(Self::Text(Arc::from(text)));
229 }
230
231 // Check for wrapped Text variant: { "text": "..." }
232 if let Some((_key, value)) = obj.iter().find(|(k, _)| k.to_lowercase() == "text")
233 && obj.len() == 1
234 {
235 // Only one field, and it's "text" (case-insensitive)
236 if let Some(text) = value.as_str() {
237 return Ok(Self::Text(Arc::from(text)));
238 }
239 }
240
241 // Check for wrapped Image variant: { "image": { "source": "...", "size": ... } }
242 if let Some((_key, value)) = obj.iter().find(|(k, _)| k.to_lowercase() == "image")
243 && obj.len() == 1
244 {
245 // Only one field, and it's "image" (case-insensitive)
246 // Try to parse the nested image object
247 if let Some(image_obj) = value.as_object()
248 && let Some(image) = LanguageModelImage::from_json(image_obj)
249 {
250 return Ok(Self::Image(image));
251 }
252 }
253
254 // Try as direct Image (object with "source" and "size" fields)
255 if let Some(image) = LanguageModelImage::from_json(obj) {
256 return Ok(Self::Image(image));
257 }
258 }
259
260 // If none of the variants match, return an error with the problematic JSON
261 Err(D::Error::custom(format!(
262 "data did not match any variant of LanguageModelToolResultContent. Expected either a string, \
263 an object with 'type': 'text', a wrapped variant like {{\"Text\": \"...\"}}, or an image object. Got: {}",
264 serde_json::to_string_pretty(&value).unwrap_or_else(|_| value.to_string())
265 )))
266 }
267}
268
269impl LanguageModelToolResultContent {
270 pub fn to_str(&self) -> Option<&str> {
271 match self {
272 Self::Text(text) => Some(text),
273 Self::Image(_) => None,
274 }
275 }
276
277 pub fn is_empty(&self) -> bool {
278 match self {
279 Self::Text(text) => text.chars().all(|c| c.is_whitespace()),
280 Self::Image(_) => false,
281 }
282 }
283}
284
285impl From<&str> for LanguageModelToolResultContent {
286 fn from(value: &str) -> Self {
287 Self::Text(Arc::from(value))
288 }
289}
290
291impl From<String> for LanguageModelToolResultContent {
292 fn from(value: String) -> Self {
293 Self::Text(Arc::from(value))
294 }
295}
296
297impl From<LanguageModelImage> for LanguageModelToolResultContent {
298 fn from(image: LanguageModelImage) -> Self {
299 Self::Image(image)
300 }
301}
302
303#[derive(Debug, Clone, Serialize, Deserialize, Eq, PartialEq, Hash)]
304pub enum MessageContent {
305 Text(String),
306 Thinking {
307 text: String,
308 signature: Option<String>,
309 },
310 RedactedThinking(String),
311 Image(LanguageModelImage),
312 ToolUse(LanguageModelToolUse),
313 ToolResult(LanguageModelToolResult),
314}
315
316impl MessageContent {
317 pub fn to_str(&self) -> Option<&str> {
318 match self {
319 MessageContent::Text(text) => Some(text.as_str()),
320 MessageContent::Thinking { text, .. } => Some(text.as_str()),
321 MessageContent::RedactedThinking(_) => None,
322 MessageContent::ToolResult(tool_result) => tool_result.content.to_str(),
323 MessageContent::ToolUse(_) | MessageContent::Image(_) => None,
324 }
325 }
326
327 pub fn is_empty(&self) -> bool {
328 match self {
329 MessageContent::Text(text) => text.chars().all(|c| c.is_whitespace()),
330 MessageContent::Thinking { text, .. } => text.chars().all(|c| c.is_whitespace()),
331 MessageContent::ToolResult(tool_result) => tool_result.content.is_empty(),
332 MessageContent::RedactedThinking(_)
333 | MessageContent::ToolUse(_)
334 | MessageContent::Image(_) => false,
335 }
336 }
337}
338
339impl From<String> for MessageContent {
340 fn from(value: String) -> Self {
341 MessageContent::Text(value)
342 }
343}
344
345impl From<&str> for MessageContent {
346 fn from(value: &str) -> Self {
347 MessageContent::Text(value.to_string())
348 }
349}
350
351#[derive(Clone, Serialize, Deserialize, Debug, PartialEq, Hash)]
352pub struct LanguageModelRequestMessage {
353 pub role: Role,
354 pub content: Vec<MessageContent>,
355 pub cache: bool,
356}
357
358impl LanguageModelRequestMessage {
359 pub fn string_contents(&self) -> String {
360 let mut buffer = String::new();
361 for string in self.content.iter().filter_map(|content| content.to_str()) {
362 buffer.push_str(string);
363 }
364
365 buffer
366 }
367
368 pub fn contents_empty(&self) -> bool {
369 self.content.iter().all(|content| content.is_empty())
370 }
371}
372
373#[derive(Debug, PartialEq, Hash, Clone, Serialize, Deserialize)]
374pub struct LanguageModelRequestTool {
375 pub name: String,
376 pub description: String,
377 pub input_schema: serde_json::Value,
378}
379
380#[derive(Debug, PartialEq, Hash, Clone, Serialize, Deserialize)]
381pub enum LanguageModelToolChoice {
382 Auto,
383 Any,
384 None,
385}
386
387#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq)]
388pub struct LanguageModelRequest {
389 pub thread_id: Option<String>,
390 pub prompt_id: Option<String>,
391 pub intent: Option<CompletionIntent>,
392 pub mode: Option<CompletionMode>,
393 pub messages: Vec<LanguageModelRequestMessage>,
394 pub tools: Vec<LanguageModelRequestTool>,
395 pub tool_choice: Option<LanguageModelToolChoice>,
396 pub stop: Vec<String>,
397 pub temperature: Option<f32>,
398 pub thinking_allowed: bool,
399}
400
401#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
402pub struct LanguageModelResponseMessage {
403 pub role: Option<Role>,
404 pub content: Option<String>,
405}
406
407#[cfg(test)]
408mod tests {
409 use super::*;
410
411 #[test]
412 fn test_language_model_tool_result_content_deserialization() {
413 let json = r#""This is plain text""#;
414 let result: LanguageModelToolResultContent = serde_json::from_str(json).unwrap();
415 assert_eq!(
416 result,
417 LanguageModelToolResultContent::Text("This is plain text".into())
418 );
419
420 let json = r#"{"type": "text", "text": "This is wrapped text"}"#;
421 let result: LanguageModelToolResultContent = serde_json::from_str(json).unwrap();
422 assert_eq!(
423 result,
424 LanguageModelToolResultContent::Text("This is wrapped text".into())
425 );
426
427 let json = r#"{"Type": "TEXT", "TEXT": "Case insensitive"}"#;
428 let result: LanguageModelToolResultContent = serde_json::from_str(json).unwrap();
429 assert_eq!(
430 result,
431 LanguageModelToolResultContent::Text("Case insensitive".into())
432 );
433
434 let json = r#"{"Text": "Wrapped variant"}"#;
435 let result: LanguageModelToolResultContent = serde_json::from_str(json).unwrap();
436 assert_eq!(
437 result,
438 LanguageModelToolResultContent::Text("Wrapped variant".into())
439 );
440
441 let json = r#"{"text": "Lowercase wrapped"}"#;
442 let result: LanguageModelToolResultContent = serde_json::from_str(json).unwrap();
443 assert_eq!(
444 result,
445 LanguageModelToolResultContent::Text("Lowercase wrapped".into())
446 );
447
448 // Test image deserialization
449 let json = r#"{
450 "source": "base64encodedimagedata",
451 "size": {
452 "width": 100,
453 "height": 200
454 }
455 }"#;
456 let result: LanguageModelToolResultContent = serde_json::from_str(json).unwrap();
457 match result {
458 LanguageModelToolResultContent::Image(image) => {
459 assert_eq!(image.source.as_ref(), "base64encodedimagedata");
460 assert_eq!(image.size.width.0, 100);
461 assert_eq!(image.size.height.0, 200);
462 }
463 _ => panic!("Expected Image variant"),
464 }
465
466 // Test wrapped Image variant
467 let json = r#"{
468 "Image": {
469 "source": "wrappedimagedata",
470 "size": {
471 "width": 50,
472 "height": 75
473 }
474 }
475 }"#;
476 let result: LanguageModelToolResultContent = serde_json::from_str(json).unwrap();
477 match result {
478 LanguageModelToolResultContent::Image(image) => {
479 assert_eq!(image.source.as_ref(), "wrappedimagedata");
480 assert_eq!(image.size.width.0, 50);
481 assert_eq!(image.size.height.0, 75);
482 }
483 _ => panic!("Expected Image variant"),
484 }
485
486 // Test wrapped Image variant with case insensitive
487 let json = r#"{
488 "image": {
489 "Source": "caseinsensitive",
490 "SIZE": {
491 "width": 30,
492 "height": 40
493 }
494 }
495 }"#;
496 let result: LanguageModelToolResultContent = serde_json::from_str(json).unwrap();
497 match result {
498 LanguageModelToolResultContent::Image(image) => {
499 assert_eq!(image.source.as_ref(), "caseinsensitive");
500 assert_eq!(image.size.width.0, 30);
501 assert_eq!(image.size.height.0, 40);
502 }
503 _ => panic!("Expected Image variant"),
504 }
505
506 // Test that wrapped text with wrong type fails
507 let json = r#"{"type": "blahblah", "text": "This should fail"}"#;
508 let result: Result<LanguageModelToolResultContent, _> = serde_json::from_str(json);
509 assert!(result.is_err());
510
511 // Test that malformed JSON fails
512 let json = r#"{"invalid": "structure"}"#;
513 let result: Result<LanguageModelToolResultContent, _> = serde_json::from_str(json);
514 assert!(result.is_err());
515
516 // Test edge cases
517 let json = r#""""#; // Empty string
518 let result: LanguageModelToolResultContent = serde_json::from_str(json).unwrap();
519 assert_eq!(result, LanguageModelToolResultContent::Text("".into()));
520
521 // Test with extra fields in wrapped text (should be ignored)
522 let json = r#"{"type": "text", "text": "Hello", "extra": "field"}"#;
523 let result: LanguageModelToolResultContent = serde_json::from_str(json).unwrap();
524 assert_eq!(result, LanguageModelToolResultContent::Text("Hello".into()));
525
526 // Test direct image with case-insensitive fields
527 let json = r#"{
528 "SOURCE": "directimage",
529 "Size": {
530 "width": 200,
531 "height": 300
532 }
533 }"#;
534 let result: LanguageModelToolResultContent = serde_json::from_str(json).unwrap();
535 match result {
536 LanguageModelToolResultContent::Image(image) => {
537 assert_eq!(image.source.as_ref(), "directimage");
538 assert_eq!(image.size.width.0, 200);
539 assert_eq!(image.size.height.0, 300);
540 }
541 _ => panic!("Expected Image variant"),
542 }
543
544 // Test that multiple fields prevent wrapped variant interpretation
545 let json = r#"{"Text": "not wrapped", "extra": "field"}"#;
546 let result: Result<LanguageModelToolResultContent, _> = serde_json::from_str(json);
547 assert!(result.is_err());
548
549 // Test wrapped text with uppercase TEXT variant
550 let json = r#"{"TEXT": "Uppercase variant"}"#;
551 let result: LanguageModelToolResultContent = serde_json::from_str(json).unwrap();
552 assert_eq!(
553 result,
554 LanguageModelToolResultContent::Text("Uppercase variant".into())
555 );
556
557 // Test that numbers and other JSON values fail gracefully
558 let json = r#"123"#;
559 let result: Result<LanguageModelToolResultContent, _> = serde_json::from_str(json);
560 assert!(result.is_err());
561
562 let json = r#"null"#;
563 let result: Result<LanguageModelToolResultContent, _> = serde_json::from_str(json);
564 assert!(result.is_err());
565
566 let json = r#"[1, 2, 3]"#;
567 let result: Result<LanguageModelToolResultContent, _> = serde_json::from_str(json);
568 assert!(result.is_err());
569 }
570}