1use crate::{
2 PredictionProvider,
3 example::{ActualCursor, Example},
4 format_prompt::TeacherPrompt,
5 repair,
6};
7use anyhow::{Context as _, Result};
8use zeta_prompt::{CURSOR_MARKER, ZetaFormat};
9
10pub fn run_parse_output(example: &mut Example) -> Result<()> {
11 example
12 .prompt_inputs
13 .as_ref()
14 .context("prompt_inputs required")?;
15
16 let to_parse: Vec<_> = example
17 .predictions
18 .iter()
19 .enumerate()
20 .filter(|(_, p)| !p.actual_output.is_empty())
21 .map(|(ix, p)| (ix, p.actual_output.clone(), p.provider))
22 .collect();
23
24 for (ix, actual_output, provider) in to_parse {
25 let (actual_patch, actual_cursor) =
26 parse_prediction_output(example, &actual_output, provider)?;
27 example.predictions[ix].actual_patch = Some(actual_patch);
28 example.predictions[ix].actual_cursor = actual_cursor;
29 }
30
31 Ok(())
32}
33
34pub fn parse_prediction_output(
35 example: &Example,
36 actual_output: &str,
37 provider: PredictionProvider,
38) -> Result<(String, Option<ActualCursor>)> {
39 match provider {
40 PredictionProvider::Teacher(_) | PredictionProvider::TeacherNonBatching(_) => {
41 TeacherPrompt::parse(example, actual_output)
42 }
43 PredictionProvider::Zeta2(version) => parse_zeta2_output(example, actual_output, version),
44 PredictionProvider::Repair => repair::parse(example, actual_output),
45 _ => anyhow::bail!(
46 "parse-output only supports Teacher and Zeta2 providers, got {:?}",
47 provider
48 ),
49 }
50}
51
52fn extract_zeta2_current_region(prompt: &str, format: ZetaFormat) -> Result<String> {
53 let (current_marker, end_marker) = match format {
54 ZetaFormat::V0112MiddleAtEnd => ("<|fim_middle|>current\n", "<|fim_middle|>updated"),
55 ZetaFormat::V0113Ordered | ZetaFormat::V0114180EditableRegion => {
56 ("<|fim_middle|>current\n", "<|fim_suffix|>")
57 }
58 ZetaFormat::V0120GitMergeMarkers
59 | ZetaFormat::V0131GitMergeMarkersPrefix
60 | ZetaFormat::V0211Prefill => (
61 zeta_prompt::v0120_git_merge_markers::START_MARKER,
62 zeta_prompt::v0120_git_merge_markers::SEPARATOR,
63 ),
64 };
65
66 let start = prompt.find(current_marker).with_context(|| {
67 format!(
68 "missing current marker '{}' in prompt",
69 current_marker.trim()
70 )
71 })? + current_marker.len();
72
73 let end = prompt[start..]
74 .find(end_marker)
75 .with_context(|| format!("missing end marker '{}' in prompt", end_marker.trim()))?
76 + start;
77
78 let region = &prompt[start..end];
79 let region = region.replace(CURSOR_MARKER, "");
80
81 Ok(region)
82}
83
84fn parse_zeta2_output(
85 example: &Example,
86 actual_output: &str,
87 format: ZetaFormat,
88) -> Result<(String, Option<ActualCursor>)> {
89 let prompt = &example.prompt.as_ref().context("prompt required")?.input;
90 let prompt_inputs = example
91 .prompt_inputs
92 .as_ref()
93 .context("prompt_inputs required")?;
94
95 let old_text = extract_zeta2_current_region(prompt, format)?;
96
97 let mut new_text = actual_output.to_string();
98 let cursor_offset = if let Some(offset) = new_text.find(CURSOR_MARKER) {
99 new_text.replace_range(offset..offset + CURSOR_MARKER.len(), "");
100 Some(offset)
101 } else {
102 None
103 };
104
105 let suffix = match format {
106 ZetaFormat::V0131GitMergeMarkersPrefix | ZetaFormat::V0211Prefill => {
107 zeta_prompt::v0131_git_merge_markers_prefix::END_MARKER
108 }
109 ZetaFormat::V0120GitMergeMarkers => zeta_prompt::v0120_git_merge_markers::END_MARKER,
110 ZetaFormat::V0112MiddleAtEnd
111 | ZetaFormat::V0113Ordered
112 | ZetaFormat::V0114180EditableRegion => "",
113 };
114 if !suffix.is_empty() {
115 new_text = new_text
116 .strip_suffix(suffix)
117 .unwrap_or(&new_text)
118 .to_string();
119 }
120
121 let mut old_text_normalized = old_text.clone();
122 if !new_text.is_empty() && !new_text.ends_with('\n') {
123 new_text.push('\n');
124 }
125 if !old_text_normalized.is_empty() && !old_text_normalized.ends_with('\n') {
126 old_text_normalized.push('\n');
127 }
128
129 let old_text_trimmed = old_text.trim_end_matches('\n');
130 let (editable_region_offset, _) = prompt_inputs
131 .content
132 .match_indices(old_text_trimmed)
133 .min_by_key(|(index, _)| index.abs_diff(prompt_inputs.cursor_offset))
134 .with_context(|| {
135 format!(
136 "could not find editable region in content.\nLooking for:\n{}\n\nIn content:\n{}",
137 old_text_trimmed, &prompt_inputs.content
138 )
139 })?;
140
141 let editable_region_start_line = prompt_inputs.content[..editable_region_offset]
142 .matches('\n')
143 .count();
144
145 // Use full context so cursor offset (relative to editable region start) aligns with diff content
146 let editable_region_lines = old_text_normalized.lines().count() as u32;
147 let diff = language::unified_diff_with_context(
148 &old_text_normalized,
149 &new_text,
150 editable_region_start_line as u32,
151 editable_region_start_line as u32,
152 editable_region_lines,
153 );
154
155 let formatted_diff = format!(
156 "--- a/{path}\n+++ b/{path}\n{diff}",
157 path = example.spec.cursor_path.to_string_lossy(),
158 );
159
160 let actual_cursor = cursor_offset.map(|editable_region_cursor_offset| {
161 ActualCursor::from_editable_region(
162 &example.spec.cursor_path,
163 editable_region_cursor_offset,
164 &new_text,
165 &prompt_inputs.content,
166 editable_region_offset,
167 editable_region_start_line,
168 )
169 });
170
171 Ok((formatted_diff, actual_cursor))
172}
173
174#[cfg(test)]
175mod tests {
176 use super::*;
177
178 #[test]
179 fn test_extract_zeta2_current_region_v0113() {
180 let prompt = indoc::indoc! {"
181 <|file_sep|>src/main.rs
182 <|fim_prefix|>
183 fn main() {
184 <|fim_middle|>current
185 println!(\"hello\");
186 <|fim_suffix|>
187 }
188 <|fim_middle|>updated
189 "};
190
191 let region = extract_zeta2_current_region(prompt, ZetaFormat::V0113Ordered).unwrap();
192 assert_eq!(region, "println!(\"hello\");\n");
193 }
194
195 #[test]
196 fn test_extract_zeta2_current_region_v0112() {
197 let prompt = indoc::indoc! {"
198 <|file_sep|>src/main.rs
199 <|fim_prefix|>
200 fn main() {
201 <|fim_suffix|>
202 }
203 <|fim_middle|>current
204 println!(\"hello\");
205 <|fim_middle|>updated
206 "};
207
208 let region = extract_zeta2_current_region(prompt, ZetaFormat::V0112MiddleAtEnd).unwrap();
209 assert_eq!(region, "println!(\"hello\");\n");
210 }
211
212 #[test]
213 fn test_extract_zeta2_current_region_with_cursor_marker() {
214 let prompt = indoc::indoc! {"
215 <|file_sep|>src/main.rs
216 <|fim_prefix|>
217 fn main() {
218 <|fim_middle|>current
219 print<|user_cursor|>ln!(\"hello\");
220 <|fim_suffix|>
221 }
222 <|fim_middle|>updated
223 "};
224
225 let region = extract_zeta2_current_region(prompt, ZetaFormat::V0113Ordered).unwrap();
226 assert_eq!(region, "println!(\"hello\");\n");
227 }
228
229 #[test]
230 fn test_extract_zeta2_current_region_v0120_git_merge_markers() {
231 let prompt = indoc::indoc! {"
232 <|file_sep|>src/main.rs
233 <|fim_prefix|>
234 fn main() {
235 <|fim_suffix|>
236 }
237 <|fim_middle|><<<<<<< CURRENT
238 println!(\"hello\");
239 =======
240 "};
241
242 let region =
243 extract_zeta2_current_region(prompt, ZetaFormat::V0120GitMergeMarkers).unwrap();
244 assert_eq!(region, "println!(\"hello\");\n");
245 }
246
247 #[test]
248 fn test_extract_zeta2_current_region_v0120_with_cursor_marker() {
249 let prompt = indoc::indoc! {"
250 <|file_sep|>src/main.rs
251 <|fim_prefix|>
252 fn main() {
253 <|fim_suffix|>
254 }
255 <|fim_middle|><<<<<<< CURRENT
256 print<|user_cursor|>ln!(\"hello\");
257 =======
258 "};
259
260 let region =
261 extract_zeta2_current_region(prompt, ZetaFormat::V0120GitMergeMarkers).unwrap();
262 assert_eq!(region, "println!(\"hello\");\n");
263 }
264}