1use crate::PredictionProvider;
2use crate::paths::WORKTREES_DIR;
3use crate::qa::QaResult;
4use anyhow::{Context as _, Result};
5use collections::HashMap;
6use edit_prediction::example_spec::ExampleSpec;
7use edit_prediction::udiff::OpenedBuffers;
8use gpui::Entity;
9use http_client::Url;
10use language::{Anchor, Buffer};
11use project::Project;
12use serde::{Deserialize, Serialize};
13use std::{
14 borrow::Cow,
15 collections::VecDeque,
16 io::Read,
17 path::{Path, PathBuf},
18 sync::Arc,
19};
20use zeta_prompt::RelatedFile;
21
22#[derive(Clone, Debug, Serialize, Deserialize)]
23pub struct Example {
24 #[serde(flatten)]
25 pub spec: ExampleSpec,
26
27 /// The full content of the file where an edit is being predicted, and the
28 /// actual cursor offset.
29 #[serde(skip_serializing_if = "Option::is_none")]
30 pub prompt_inputs: Option<ExamplePromptInputs>,
31
32 /// The input and expected output from the edit prediction model.
33 #[serde(skip_serializing_if = "Option::is_none")]
34 pub prompt: Option<ExamplePrompt>,
35
36 /// The actual predictions from the model.
37 #[serde(default, skip_serializing_if = "Vec::is_empty")]
38 pub predictions: Vec<ExamplePrediction>,
39
40 /// The scores, for how well the actual predictions match the expected
41 /// predictions.
42 #[serde(default, skip_serializing_if = "Vec::is_empty")]
43 pub score: Vec<ExampleScore>,
44
45 /// QA evaluation results for each prediction (indexed parallel to `predictions`).
46 #[serde(default, skip_serializing_if = "Vec::is_empty")]
47 pub qa: Vec<Option<QaResult>>,
48
49 /// The application state used to process this example.
50 #[serde(skip)]
51 pub state: Option<ExampleState>,
52}
53
54#[derive(Clone, Debug)]
55pub struct ExampleState {
56 pub project: Entity<Project>,
57 pub buffer: Entity<Buffer>,
58 pub cursor_position: Anchor,
59 pub _open_buffers: OpenedBuffers,
60}
61
62#[derive(Clone, Debug, Serialize, Deserialize)]
63pub struct ExamplePromptInputs {
64 pub content: String,
65 pub cursor_row: u32,
66 pub cursor_column: u32,
67 pub cursor_offset: usize,
68 #[serde(default, skip_serializing_if = "Option::is_none")]
69 pub excerpt_start_row: Option<u32>,
70 pub edit_history: Vec<Arc<zeta_prompt::Event>>,
71 pub related_files: Option<Vec<RelatedFile>>,
72}
73
74#[derive(Clone, Debug, Serialize, Deserialize)]
75pub struct ExamplePrompt {
76 pub input: String,
77 pub expected_output: String,
78 pub rejected_output: Option<String>, // For DPO
79 pub provider: PredictionProvider,
80}
81
82#[derive(Clone, Debug, Serialize, Deserialize)]
83pub struct ExamplePrediction {
84 #[serde(default, skip_serializing_if = "Option::is_none")]
85 pub actual_patch: Option<String>,
86 #[serde(deserialize_with = "deserialize_null_as_empty_string")]
87 pub actual_output: String,
88 #[serde(default, skip_serializing_if = "Option::is_none")]
89 pub error: Option<String>,
90 pub provider: PredictionProvider,
91}
92
93fn deserialize_null_as_empty_string<'de, D>(deserializer: D) -> Result<String, D::Error>
94where
95 D: serde::Deserializer<'de>,
96{
97 let opt = Option::<String>::deserialize(deserializer)?;
98 Ok(opt.unwrap_or_default())
99}
100
101#[derive(Clone, Debug, Serialize, Deserialize)]
102pub struct ExampleScore {
103 pub delta_chr_f: f32,
104 pub braces_disbalance: usize,
105 #[serde(default)]
106 pub exact_lines_tp: usize,
107 #[serde(default)]
108 pub exact_lines_fp: usize,
109 #[serde(default)]
110 pub exact_lines_fn: usize,
111 #[serde(default)]
112 pub reversal_ratio: f32,
113}
114
115impl Example {
116 pub fn repo_name(&self) -> Result<RepoName<'_>> {
117 // git@github.com:owner/repo.git
118 if self.spec.repository_url.contains('@') {
119 let (owner, repo) = self
120 .spec
121 .repository_url
122 .split_once(':')
123 .context("expected : in git url")?
124 .1
125 .split_once('/')
126 .context("expected / in git url")?;
127 Ok(RepoName {
128 owner: Cow::Borrowed(owner),
129 name: Cow::Borrowed(repo.trim_end_matches(".git")),
130 })
131 // http://github.com/owner/repo.git
132 } else {
133 let url = Url::parse(&self.spec.repository_url)?;
134 let mut segments = url.path_segments().context("empty http url")?;
135 let owner = segments
136 .next()
137 .context("expected owner path segment")?
138 .to_string();
139 let repo = segments
140 .next()
141 .context("expected repo path segment")?
142 .trim_end_matches(".git")
143 .to_string();
144 assert!(segments.next().is_none());
145
146 Ok(RepoName {
147 owner: Cow::Owned(owner),
148 name: Cow::Owned(repo),
149 })
150 }
151 }
152}
153
154pub struct RepoName<'a> {
155 pub owner: Cow<'a, str>,
156 pub name: Cow<'a, str>,
157}
158
159impl RepoName<'_> {
160 pub fn worktree_path(&self) -> PathBuf {
161 WORKTREES_DIR
162 .join(self.owner.as_ref())
163 .join(self.name.as_ref())
164 }
165}
166
167pub fn read_example_files(inputs: &[PathBuf]) -> Vec<Example> {
168 let mut examples = Vec::new();
169
170 for path in inputs {
171 let is_stdin = path.as_path() == Path::new("-");
172 let content = if is_stdin {
173 let mut buffer = String::new();
174 std::io::stdin()
175 .read_to_string(&mut buffer)
176 .expect("Failed to read from stdin");
177 buffer
178 } else {
179 std::fs::read_to_string(path)
180 .unwrap_or_else(|_| panic!("Failed to read path: {:?}", &path))
181 };
182 let filename = path.file_stem().unwrap().to_string_lossy().to_string();
183 let ext = if !is_stdin {
184 path.extension()
185 .map(|ext| ext.to_string_lossy().to_string())
186 .unwrap_or_else(|| panic!("{} should have an extension", path.display()))
187 } else {
188 "jsonl".to_string()
189 };
190
191 match ext.as_ref() {
192 "json" => {
193 let mut example =
194 serde_json::from_str::<Example>(&content).unwrap_or_else(|error| {
195 panic!("Failed to parse example file: {}\n{error}", path.display())
196 });
197 if example.spec.name.is_empty() {
198 example.spec.name = filename;
199 }
200 examples.push(example);
201 }
202 "jsonl" => examples.extend(
203 content
204 .lines()
205 .enumerate()
206 .map(|(line_ix, line)| {
207 let mut example =
208 serde_json::from_str::<Example>(line).unwrap_or_else(|error| {
209 panic!(
210 "Failed to parse example on {}:{}\n{error}",
211 path.display(),
212 line_ix + 1
213 )
214 });
215 if example.spec.name.is_empty() {
216 example.spec.name = format!("{filename}-{line_ix}")
217 }
218 example
219 })
220 .collect::<Vec<Example>>(),
221 ),
222 "md" => {
223 let mut example = parse_markdown_example(&content).unwrap();
224 if example.spec.name.is_empty() {
225 example.spec.name = filename;
226 }
227 examples.push(example);
228 }
229 ext => {
230 panic!("{} has invalid example extension `{ext}`", path.display())
231 }
232 }
233 }
234
235 examples
236}
237
238pub fn sort_examples_by_repo_and_rev(examples: &mut [Example]) {
239 examples.sort_by(|a, b| {
240 a.spec
241 .repository_url
242 .cmp(&b.spec.repository_url)
243 .then(b.spec.revision.cmp(&a.spec.revision))
244 });
245}
246
247pub fn group_examples_by_repo(examples: Vec<Example>) -> VecDeque<Vec<Example>> {
248 let mut examples_by_repo: HashMap<String, Vec<Example>> = HashMap::default();
249 let mut ungrouped = Vec::new();
250 for example in examples {
251 if example.spec.repository_url.is_empty() {
252 ungrouped.push(example);
253 } else {
254 examples_by_repo
255 .entry(example.spec.repository_url.clone())
256 .or_insert_with(Vec::new)
257 .push(example);
258 }
259 }
260 let mut result: VecDeque<Vec<Example>> = examples_by_repo.into_values().collect();
261 for example in ungrouped {
262 result.push_back(vec![example]);
263 }
264 result
265}
266
267fn parse_markdown_example(input: &str) -> Result<Example> {
268 let spec = ExampleSpec::from_markdown(input)?;
269 Ok(Example {
270 spec,
271 prompt_inputs: None,
272 prompt: None,
273 predictions: Vec::new(),
274 score: Vec::new(),
275 qa: Vec::new(),
276 state: None,
277 })
278}