1use crate::{
2 PromptFormat,
3 example::{Example, ExamplePrompt},
4 headless::EpAppState,
5 load_project::run_load_project,
6 progress::{Progress, Step},
7 retrieve_context::run_context_retrieval,
8};
9use anyhow::{Context as _, Result, ensure};
10use edit_prediction::{
11 EditPredictionStore,
12 zeta2::{zeta2_output_for_patch, zeta2_prompt_input},
13};
14use gpui::AsyncApp;
15use std::sync::Arc;
16use zeta_prompt::format_zeta_prompt;
17
18pub async fn run_format_prompt(
19 example: &mut Example,
20 prompt_format: PromptFormat,
21 app_state: Arc<EpAppState>,
22 mut cx: AsyncApp,
23) -> Result<()> {
24 run_context_retrieval(example, app_state.clone(), cx.clone()).await?;
25
26 let _step_progress = Progress::global().start(Step::FormatPrompt, &example.spec.name);
27
28 match prompt_format {
29 PromptFormat::Teacher => {
30 let prompt = TeacherPrompt::format_prompt(example);
31 example.prompt = Some(ExamplePrompt {
32 input: prompt,
33 expected_output: example.spec.expected_patch.clone(), // TODO
34 format: prompt_format,
35 });
36 }
37 PromptFormat::Zeta2 => {
38 run_load_project(example, app_state, cx.clone()).await?;
39
40 let ep_store = cx.update(|cx| {
41 EditPredictionStore::try_global(cx).context("EditPredictionStore not initialized")
42 })??;
43
44 let state = example.state.as_ref().context("state must be set")?;
45 let snapshot = state.buffer.read_with(&cx, |buffer, _| buffer.snapshot())?;
46 let project = state.project.clone();
47 let (_, input) = ep_store.update(&mut cx, |ep_store, cx| {
48 let events = ep_store
49 .edit_history_for_project(&project, cx)
50 .into_iter()
51 .map(|e| e.event)
52 .collect();
53 anyhow::Ok(zeta2_prompt_input(
54 &snapshot,
55 example
56 .context
57 .as_ref()
58 .context("context must be set")?
59 .files
60 .clone(),
61 events,
62 example.spec.cursor_path.clone(),
63 example
64 .buffer
65 .as_ref()
66 .context("buffer must be set")?
67 .cursor_offset,
68 ))
69 })??;
70 let prompt = format_zeta_prompt(&input);
71 let expected_output =
72 zeta2_output_for_patch(&input, &example.spec.expected_patch.clone())?;
73 example.prompt = Some(ExamplePrompt {
74 input: prompt,
75 expected_output,
76 format: prompt_format,
77 });
78 }
79 };
80 Ok(())
81}
82
83pub struct TeacherPrompt;
84
85impl TeacherPrompt {
86 const PROMPT: &str = include_str!("teacher.prompt.md");
87 pub(crate) const EDITABLE_REGION_START: &str = "<|editable_region_start|>\n";
88 pub(crate) const EDITABLE_REGION_END: &str = "<|editable_region_end|>";
89
90 /// Truncate edit history to this number of last lines
91 const MAX_HISTORY_LINES: usize = 128;
92
93 pub fn format_prompt(example: &Example) -> String {
94 let edit_history = Self::format_edit_history(&example.spec.edit_history);
95 let context = Self::format_context(example);
96 let editable_region = Self::format_editable_region(example);
97
98 let prompt = Self::PROMPT
99 .replace("{{context}}", &context)
100 .replace("{{edit_history}}", &edit_history)
101 .replace("{{editable_region}}", &editable_region);
102
103 prompt
104 }
105
106 pub fn parse(example: &Example, response: &str) -> Result<String> {
107 // Ideally, we should always be able to find cursor position in the retrieved context.
108 // In reality, sometimes we don't find it for these reasons:
109 // 1. `example.cursor_position` contains _more_ context than included in the retrieved context
110 // (can be fixed by getting cursor coordinates at the load_example stage)
111 // 2. Context retriever just didn't include cursor line.
112 //
113 // In that case, fallback to using `cursor_position` as excerpt.
114 let cursor_file = &example
115 .buffer
116 .as_ref()
117 .context("`buffer` should be filled in in the context collection step")?
118 .content;
119
120 // Extract updated (new) editable region from the model response
121 let new_editable_region = extract_last_codeblock(response);
122
123 // Reconstruct old editable region we sent to the model
124 let old_editable_region = Self::format_editable_region(example);
125 let old_editable_region = Self::extract_editable_region(&old_editable_region);
126 ensure!(
127 cursor_file.contains(&old_editable_region),
128 "Something's wrong: editable_region is not found in the cursor file"
129 );
130
131 // Apply editable region to a larger context and compute diff.
132 // This is needed to get a better context lines around the editable region
133 let edited_file = cursor_file.replace(&old_editable_region, &new_editable_region);
134 let diff = language::unified_diff(&cursor_file, &edited_file);
135
136 let diff = indoc::formatdoc! {"
137 --- a/{path}
138 +++ b/{path}
139 {diff}",
140 path = example.spec.cursor_path.to_string_lossy(),
141 diff = diff,
142 };
143
144 Ok(diff)
145 }
146
147 fn format_edit_history(edit_history: &str) -> String {
148 // Strip comments ("garbage lines") from edit history
149 let lines = edit_history
150 .lines()
151 .filter(|&s| Self::is_udiff_content_line(s))
152 .collect::<Vec<_>>();
153
154 let history_lines = if lines.len() > Self::MAX_HISTORY_LINES {
155 &lines[lines.len() - Self::MAX_HISTORY_LINES..]
156 } else {
157 &lines
158 };
159
160 if history_lines.is_empty() {
161 return "(No edit history)".to_string();
162 }
163
164 history_lines.join("\n")
165 }
166
167 fn format_context(example: &Example) -> String {
168 assert!(example.context.is_some(), "Missing context retriever step");
169
170 let mut prompt = String::new();
171 zeta_prompt::write_related_files(&mut prompt, &example.context.as_ref().unwrap().files);
172
173 prompt
174 }
175
176 fn format_editable_region(example: &Example) -> String {
177 let mut result = String::new();
178
179 let path_str = example.spec.cursor_path.to_string_lossy();
180 result.push_str(&format!("`````path=\"{path_str}\"\n"));
181 result.push_str(Self::EDITABLE_REGION_START);
182
183 // TODO: control number of lines around cursor
184 result.push_str(&example.spec.cursor_position);
185 if !example.spec.cursor_position.ends_with('\n') {
186 result.push('\n');
187 }
188
189 result.push_str(&format!("{}\n", Self::EDITABLE_REGION_END));
190 result.push_str("`````");
191
192 result
193 }
194
195 fn extract_editable_region(text: &str) -> String {
196 let start = text
197 .find(Self::EDITABLE_REGION_START)
198 .map_or(0, |pos| pos + Self::EDITABLE_REGION_START.len());
199 let end = text.find(Self::EDITABLE_REGION_END).unwrap_or(text.len());
200
201 let region = &text[start..end];
202
203 region.replace("<|user_cursor|>", "")
204 }
205
206 fn is_udiff_content_line(s: &str) -> bool {
207 s.starts_with("-")
208 || s.starts_with("+")
209 || s.starts_with(" ")
210 || s.starts_with("---")
211 || s.starts_with("+++")
212 || s.starts_with("@@")
213 }
214}
215
216fn extract_last_codeblock(text: &str) -> String {
217 let mut last_block = None;
218 let mut search_start = 0;
219
220 while let Some(start) = text[search_start..].find("```") {
221 let start = start + search_start;
222 let bytes = text.as_bytes();
223 let mut backtick_end = start;
224
225 while backtick_end < bytes.len() && bytes[backtick_end] == b'`' {
226 backtick_end += 1;
227 }
228
229 let backtick_count = backtick_end - start;
230 let closing_backticks = "`".repeat(backtick_count);
231
232 while backtick_end < bytes.len() && bytes[backtick_end] != b'\n' {
233 backtick_end += 1;
234 }
235
236 if let Some(end_pos) = text[backtick_end..].find(&closing_backticks) {
237 let code_block = &text[backtick_end + 1..backtick_end + end_pos];
238 last_block = Some(code_block.to_string());
239 search_start = backtick_end + end_pos + backtick_count;
240 } else {
241 break;
242 }
243 }
244
245 last_block.unwrap_or_else(|| text.to_string())
246}
247
248#[cfg(test)]
249mod tests {
250 use super::*;
251
252 #[test]
253 fn test_extract_last_code_block() {
254 let text = indoc::indoc! {"
255 Some thinking
256
257 ```
258 first block
259 ```
260
261 `````path='something' lines=1:2
262 last block
263 `````
264 "};
265 let last_block = extract_last_codeblock(text);
266 assert_eq!(last_block, "last block\n");
267 }
268
269 #[test]
270 fn test_extract_editable_region() {
271 let text = indoc::indoc! {"
272 some lines
273 are
274 here
275 <|editable_region_start|>
276 one
277 two three
278
279 <|editable_region_end|>
280 more
281 lines here
282 "};
283 let parsed = TeacherPrompt::extract_editable_region(text);
284 assert_eq!(
285 parsed,
286 indoc::indoc! {"
287 one
288 two three
289
290 "}
291 );
292 }
293}