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 {
225 if let Some(type_str) = type_value.as_str() {
226 if type_str.to_lowercase() == "text" {
227 if let Some(text) = text_value.as_str() {
228 return Ok(Self::Text(Arc::from(text)));
229 }
230 }
231 }
232 }
233
234 // Check for wrapped Text variant: { "text": "..." }
235 if let Some((_key, value)) = obj.iter().find(|(k, _)| k.to_lowercase() == "text") {
236 if obj.len() == 1 {
237 // Only one field, and it's "text" (case-insensitive)
238 if let Some(text) = value.as_str() {
239 return Ok(Self::Text(Arc::from(text)));
240 }
241 }
242 }
243
244 // Check for wrapped Image variant: { "image": { "source": "...", "size": ... } }
245 if let Some((_key, value)) = obj.iter().find(|(k, _)| k.to_lowercase() == "image") {
246 if obj.len() == 1 {
247 // Only one field, and it's "image" (case-insensitive)
248 // Try to parse the nested image object
249 if let Some(image_obj) = value.as_object() {
250 if let Some(image) = LanguageModelImage::from_json(image_obj) {
251 return Ok(Self::Image(image));
252 }
253 }
254 }
255 }
256
257 // Try as direct Image (object with "source" and "size" fields)
258 if let Some(image) = LanguageModelImage::from_json(&obj) {
259 return Ok(Self::Image(image));
260 }
261 }
262
263 // If none of the variants match, return an error with the problematic JSON
264 Err(D::Error::custom(format!(
265 "data did not match any variant of LanguageModelToolResultContent. Expected either a string, \
266 an object with 'type': 'text', a wrapped variant like {{\"Text\": \"...\"}}, or an image object. Got: {}",
267 serde_json::to_string_pretty(&value).unwrap_or_else(|_| value.to_string())
268 )))
269 }
270}
271
272impl LanguageModelToolResultContent {
273 pub fn to_str(&self) -> Option<&str> {
274 match self {
275 Self::Text(text) => Some(&text),
276 Self::Image(_) => None,
277 }
278 }
279
280 pub fn is_empty(&self) -> bool {
281 match self {
282 Self::Text(text) => text.chars().all(|c| c.is_whitespace()),
283 Self::Image(_) => false,
284 }
285 }
286}
287
288impl From<&str> for LanguageModelToolResultContent {
289 fn from(value: &str) -> Self {
290 Self::Text(Arc::from(value))
291 }
292}
293
294impl From<String> for LanguageModelToolResultContent {
295 fn from(value: String) -> Self {
296 Self::Text(Arc::from(value))
297 }
298}
299
300#[derive(Debug, Clone, Serialize, Deserialize, Eq, PartialEq, Hash)]
301pub enum MessageContent {
302 Text(String),
303 Thinking {
304 text: String,
305 signature: Option<String>,
306 },
307 RedactedThinking(String),
308 Image(LanguageModelImage),
309 ToolUse(LanguageModelToolUse),
310 ToolResult(LanguageModelToolResult),
311}
312
313impl MessageContent {
314 pub fn to_str(&self) -> Option<&str> {
315 match self {
316 MessageContent::Text(text) => Some(text.as_str()),
317 MessageContent::Thinking { text, .. } => Some(text.as_str()),
318 MessageContent::RedactedThinking(_) => None,
319 MessageContent::ToolResult(tool_result) => tool_result.content.to_str(),
320 MessageContent::ToolUse(_) | MessageContent::Image(_) => None,
321 }
322 }
323
324 pub fn is_empty(&self) -> bool {
325 match self {
326 MessageContent::Text(text) => text.chars().all(|c| c.is_whitespace()),
327 MessageContent::Thinking { text, .. } => text.chars().all(|c| c.is_whitespace()),
328 MessageContent::ToolResult(tool_result) => tool_result.content.is_empty(),
329 MessageContent::RedactedThinking(_)
330 | MessageContent::ToolUse(_)
331 | MessageContent::Image(_) => false,
332 }
333 }
334}
335
336impl From<String> for MessageContent {
337 fn from(value: String) -> Self {
338 MessageContent::Text(value)
339 }
340}
341
342impl From<&str> for MessageContent {
343 fn from(value: &str) -> Self {
344 MessageContent::Text(value.to_string())
345 }
346}
347
348#[derive(Clone, Serialize, Deserialize, Debug, PartialEq, Hash)]
349pub struct LanguageModelRequestMessage {
350 pub role: Role,
351 pub content: Vec<MessageContent>,
352 pub cache: bool,
353}
354
355impl LanguageModelRequestMessage {
356 pub fn string_contents(&self) -> String {
357 let mut buffer = String::new();
358 for string in self.content.iter().filter_map(|content| content.to_str()) {
359 buffer.push_str(string);
360 }
361
362 buffer
363 }
364
365 pub fn contents_empty(&self) -> bool {
366 self.content.iter().all(|content| content.is_empty())
367 }
368}
369
370#[derive(Debug, PartialEq, Hash, Clone, Serialize, Deserialize)]
371pub struct LanguageModelRequestTool {
372 pub name: String,
373 pub description: String,
374 pub input_schema: serde_json::Value,
375}
376
377#[derive(Debug, PartialEq, Hash, Clone, Serialize, Deserialize)]
378pub enum LanguageModelToolChoice {
379 Auto,
380 Any,
381 None,
382}
383
384#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq)]
385pub struct LanguageModelRequest {
386 pub thread_id: Option<String>,
387 pub prompt_id: Option<String>,
388 pub intent: Option<CompletionIntent>,
389 pub mode: Option<CompletionMode>,
390 pub messages: Vec<LanguageModelRequestMessage>,
391 pub tools: Vec<LanguageModelRequestTool>,
392 pub tool_choice: Option<LanguageModelToolChoice>,
393 pub stop: Vec<String>,
394 pub temperature: Option<f32>,
395 pub thinking_allowed: bool,
396}
397
398#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
399pub struct LanguageModelResponseMessage {
400 pub role: Option<Role>,
401 pub content: Option<String>,
402}
403
404#[cfg(test)]
405mod tests {
406 use super::*;
407
408 #[test]
409 fn test_language_model_tool_result_content_deserialization() {
410 let json = r#""This is plain text""#;
411 let result: LanguageModelToolResultContent = serde_json::from_str(json).unwrap();
412 assert_eq!(
413 result,
414 LanguageModelToolResultContent::Text("This is plain text".into())
415 );
416
417 let json = r#"{"type": "text", "text": "This is wrapped text"}"#;
418 let result: LanguageModelToolResultContent = serde_json::from_str(json).unwrap();
419 assert_eq!(
420 result,
421 LanguageModelToolResultContent::Text("This is wrapped text".into())
422 );
423
424 let json = r#"{"Type": "TEXT", "TEXT": "Case insensitive"}"#;
425 let result: LanguageModelToolResultContent = serde_json::from_str(json).unwrap();
426 assert_eq!(
427 result,
428 LanguageModelToolResultContent::Text("Case insensitive".into())
429 );
430
431 let json = r#"{"Text": "Wrapped variant"}"#;
432 let result: LanguageModelToolResultContent = serde_json::from_str(json).unwrap();
433 assert_eq!(
434 result,
435 LanguageModelToolResultContent::Text("Wrapped variant".into())
436 );
437
438 let json = r#"{"text": "Lowercase wrapped"}"#;
439 let result: LanguageModelToolResultContent = serde_json::from_str(json).unwrap();
440 assert_eq!(
441 result,
442 LanguageModelToolResultContent::Text("Lowercase wrapped".into())
443 );
444
445 // Test image deserialization
446 let json = r#"{
447 "source": "base64encodedimagedata",
448 "size": {
449 "width": 100,
450 "height": 200
451 }
452 }"#;
453 let result: LanguageModelToolResultContent = serde_json::from_str(json).unwrap();
454 match result {
455 LanguageModelToolResultContent::Image(image) => {
456 assert_eq!(image.source.as_ref(), "base64encodedimagedata");
457 assert_eq!(image.size.width.0, 100);
458 assert_eq!(image.size.height.0, 200);
459 }
460 _ => panic!("Expected Image variant"),
461 }
462
463 // Test wrapped Image variant
464 let json = r#"{
465 "Image": {
466 "source": "wrappedimagedata",
467 "size": {
468 "width": 50,
469 "height": 75
470 }
471 }
472 }"#;
473 let result: LanguageModelToolResultContent = serde_json::from_str(json).unwrap();
474 match result {
475 LanguageModelToolResultContent::Image(image) => {
476 assert_eq!(image.source.as_ref(), "wrappedimagedata");
477 assert_eq!(image.size.width.0, 50);
478 assert_eq!(image.size.height.0, 75);
479 }
480 _ => panic!("Expected Image variant"),
481 }
482
483 // Test wrapped Image variant with case insensitive
484 let json = r#"{
485 "image": {
486 "Source": "caseinsensitive",
487 "SIZE": {
488 "width": 30,
489 "height": 40
490 }
491 }
492 }"#;
493 let result: LanguageModelToolResultContent = serde_json::from_str(json).unwrap();
494 match result {
495 LanguageModelToolResultContent::Image(image) => {
496 assert_eq!(image.source.as_ref(), "caseinsensitive");
497 assert_eq!(image.size.width.0, 30);
498 assert_eq!(image.size.height.0, 40);
499 }
500 _ => panic!("Expected Image variant"),
501 }
502
503 // Test that wrapped text with wrong type fails
504 let json = r#"{"type": "blahblah", "text": "This should fail"}"#;
505 let result: Result<LanguageModelToolResultContent, _> = serde_json::from_str(json);
506 assert!(result.is_err());
507
508 // Test that malformed JSON fails
509 let json = r#"{"invalid": "structure"}"#;
510 let result: Result<LanguageModelToolResultContent, _> = serde_json::from_str(json);
511 assert!(result.is_err());
512
513 // Test edge cases
514 let json = r#""""#; // Empty string
515 let result: LanguageModelToolResultContent = serde_json::from_str(json).unwrap();
516 assert_eq!(result, LanguageModelToolResultContent::Text("".into()));
517
518 // Test with extra fields in wrapped text (should be ignored)
519 let json = r#"{"type": "text", "text": "Hello", "extra": "field"}"#;
520 let result: LanguageModelToolResultContent = serde_json::from_str(json).unwrap();
521 assert_eq!(result, LanguageModelToolResultContent::Text("Hello".into()));
522
523 // Test direct image with case-insensitive fields
524 let json = r#"{
525 "SOURCE": "directimage",
526 "Size": {
527 "width": 200,
528 "height": 300
529 }
530 }"#;
531 let result: LanguageModelToolResultContent = serde_json::from_str(json).unwrap();
532 match result {
533 LanguageModelToolResultContent::Image(image) => {
534 assert_eq!(image.source.as_ref(), "directimage");
535 assert_eq!(image.size.width.0, 200);
536 assert_eq!(image.size.height.0, 300);
537 }
538 _ => panic!("Expected Image variant"),
539 }
540
541 // Test that multiple fields prevent wrapped variant interpretation
542 let json = r#"{"Text": "not wrapped", "extra": "field"}"#;
543 let result: Result<LanguageModelToolResultContent, _> = serde_json::from_str(json);
544 assert!(result.is_err());
545
546 // Test wrapped text with uppercase TEXT variant
547 let json = r#"{"TEXT": "Uppercase variant"}"#;
548 let result: LanguageModelToolResultContent = serde_json::from_str(json).unwrap();
549 assert_eq!(
550 result,
551 LanguageModelToolResultContent::Text("Uppercase variant".into())
552 );
553
554 // Test that numbers and other JSON values fail gracefully
555 let json = r#"123"#;
556 let result: Result<LanguageModelToolResultContent, _> = serde_json::from_str(json);
557 assert!(result.is_err());
558
559 let json = r#"null"#;
560 let result: Result<LanguageModelToolResultContent, _> = serde_json::from_str(json);
561 assert!(result.is_err());
562
563 let json = r#"[1, 2, 3]"#;
564 let result: Result<LanguageModelToolResultContent, _> = serde_json::from_str(json);
565 assert!(result.is_err());
566 }
567}