1use anyhow::{anyhow, Result};
2use assistant_slash_command::{
3 ArgumentCompletion, SlashCommand, SlashCommandOutput, SlashCommandOutputSection,
4 SlashCommandResult,
5};
6use feature_flags::FeatureFlag;
7use futures::StreamExt;
8use gpui::{AppContext, AsyncAppContext, Task, WeakView};
9use language::{CodeLabel, LspAdapterDelegate};
10use language_model::{
11 LanguageModelCompletionEvent, LanguageModelRegistry, LanguageModelRequest,
12 LanguageModelRequestMessage, Role,
13};
14use semantic_index::{FileSummary, SemanticDb};
15use smol::channel;
16use std::sync::{atomic::AtomicBool, Arc};
17use ui::{BorrowAppContext, WindowContext};
18use util::ResultExt;
19use workspace::Workspace;
20
21use crate::slash_command::create_label_for_command;
22
23pub struct AutoSlashCommandFeatureFlag;
24
25impl FeatureFlag for AutoSlashCommandFeatureFlag {
26 const NAME: &'static str = "auto-slash-command";
27}
28
29pub(crate) struct AutoCommand;
30
31impl SlashCommand for AutoCommand {
32 fn name(&self) -> String {
33 "auto".into()
34 }
35
36 fn description(&self) -> String {
37 "Automatically infer what context to add".into()
38 }
39
40 fn menu_text(&self) -> String {
41 self.description()
42 }
43
44 fn label(&self, cx: &AppContext) -> CodeLabel {
45 create_label_for_command("auto", &["--prompt"], cx)
46 }
47
48 fn complete_argument(
49 self: Arc<Self>,
50 _arguments: &[String],
51 _cancel: Arc<AtomicBool>,
52 workspace: Option<WeakView<Workspace>>,
53 cx: &mut WindowContext,
54 ) -> Task<Result<Vec<ArgumentCompletion>>> {
55 // There's no autocomplete for a prompt, since it's arbitrary text.
56 // However, we can use this opportunity to kick off a drain of the backlog.
57 // That way, it can hopefully be done resummarizing by the time we've actually
58 // typed out our prompt. This re-runs on every keystroke during autocomplete,
59 // but in the future, we could instead do it only once, when /auto is first entered.
60 let Some(workspace) = workspace.and_then(|ws| ws.upgrade()) else {
61 log::warn!("workspace was dropped or unavailable during /auto autocomplete");
62
63 return Task::ready(Ok(Vec::new()));
64 };
65
66 let project = workspace.read(cx).project().clone();
67 let Some(project_index) =
68 cx.update_global(|index: &mut SemanticDb, cx| index.project_index(project, cx))
69 else {
70 return Task::ready(Err(anyhow!("No project indexer, cannot use /auto")));
71 };
72
73 let cx: &mut AppContext = cx;
74
75 cx.spawn(|cx: gpui::AsyncAppContext| async move {
76 let task = project_index.read_with(&cx, |project_index, cx| {
77 project_index.flush_summary_backlogs(cx)
78 })?;
79
80 cx.background_executor().spawn(task).await;
81
82 anyhow::Ok(Vec::new())
83 })
84 }
85
86 fn requires_argument(&self) -> bool {
87 true
88 }
89
90 fn run(
91 self: Arc<Self>,
92 arguments: &[String],
93 _context_slash_command_output_sections: &[SlashCommandOutputSection<language::Anchor>],
94 _context_buffer: language::BufferSnapshot,
95 workspace: WeakView<Workspace>,
96 _delegate: Option<Arc<dyn LspAdapterDelegate>>,
97 cx: &mut WindowContext,
98 ) -> Task<SlashCommandResult> {
99 let Some(workspace) = workspace.upgrade() else {
100 return Task::ready(Err(anyhow::anyhow!("workspace was dropped")));
101 };
102 if arguments.is_empty() {
103 return Task::ready(Err(anyhow!("missing prompt")));
104 };
105 let argument = arguments.join(" ");
106 let original_prompt = argument.to_string();
107 let project = workspace.read(cx).project().clone();
108 let Some(project_index) =
109 cx.update_global(|index: &mut SemanticDb, cx| index.project_index(project, cx))
110 else {
111 return Task::ready(Err(anyhow!("no project indexer")));
112 };
113
114 let task = cx.spawn(|cx: gpui::AsyncWindowContext| async move {
115 let summaries = project_index
116 .read_with(&cx, |project_index, cx| project_index.all_summaries(cx))?
117 .await?;
118
119 commands_for_summaries(&summaries, &original_prompt, &cx).await
120 });
121
122 // As a convenience, append /auto's argument to the end of the prompt
123 // so you don't have to write it again.
124 let original_prompt = argument.to_string();
125
126 cx.background_executor().spawn(async move {
127 let commands = task.await?;
128 let mut prompt = String::new();
129
130 log::info!(
131 "Translating this response into slash-commands: {:?}",
132 commands
133 );
134
135 for command in commands {
136 prompt.push('/');
137 prompt.push_str(&command.name);
138 prompt.push(' ');
139 prompt.push_str(&command.arg);
140 prompt.push('\n');
141 }
142
143 prompt.push('\n');
144 prompt.push_str(&original_prompt);
145
146 Ok(SlashCommandOutput {
147 text: prompt,
148 sections: Vec::new(),
149 run_commands_in_text: true,
150 }
151 .to_event_stream())
152 })
153 }
154}
155
156const PROMPT_INSTRUCTIONS_BEFORE_SUMMARY: &str = include_str!("prompt_before_summary.txt");
157const PROMPT_INSTRUCTIONS_AFTER_SUMMARY: &str = include_str!("prompt_after_summary.txt");
158
159fn summaries_prompt(summaries: &[FileSummary], original_prompt: &str) -> String {
160 let json_summaries = serde_json::to_string(summaries).unwrap();
161
162 format!("{PROMPT_INSTRUCTIONS_BEFORE_SUMMARY}\n{json_summaries}\n{PROMPT_INSTRUCTIONS_AFTER_SUMMARY}\n{original_prompt}")
163}
164
165/// The slash commands that the model is told about, and which we look for in the inference response.
166const SUPPORTED_SLASH_COMMANDS: &[&str] = &["search", "file"];
167
168#[derive(Debug, Clone)]
169struct CommandToRun {
170 name: String,
171 arg: String,
172}
173
174/// Given the pre-indexed file summaries for this project, as well as the original prompt
175/// string passed to `/auto`, get a list of slash commands to run, along with their arguments.
176///
177/// The prompt's output does not include the slashes (to reduce the chance that it makes a mistake),
178/// so taking one of these returned Strings and turning it into a real slash-command-with-argument
179/// involves prepending a slash to it.
180///
181/// This function will validate that each of the returned lines begins with one of SUPPORTED_SLASH_COMMANDS.
182/// Any other lines it encounters will be discarded, with a warning logged.
183async fn commands_for_summaries(
184 summaries: &[FileSummary],
185 original_prompt: &str,
186 cx: &AsyncAppContext,
187) -> Result<Vec<CommandToRun>> {
188 if summaries.is_empty() {
189 log::warn!("Inferring no context because there were no summaries available.");
190 return Ok(Vec::new());
191 }
192
193 // Use the globally configured model to translate the summaries into slash-commands,
194 // because Qwen2-7B-Instruct has not done a good job at that task.
195 let Some(model) = cx.update(|cx| LanguageModelRegistry::read_global(cx).active_model())? else {
196 log::warn!("Can't infer context because there's no active model.");
197 return Ok(Vec::new());
198 };
199 // Only go up to 90% of the actual max token count, to reduce chances of
200 // exceeding the token count due to inaccuracies in the token counting heuristic.
201 let max_token_count = (model.max_token_count() * 9) / 10;
202
203 // Rather than recursing (which would require this async function use a pinned box),
204 // we use an explicit stack of arguments and answers for when we need to "recurse."
205 let mut stack = vec![summaries];
206 let mut final_response = Vec::new();
207 let mut prompts = Vec::new();
208
209 // TODO We only need to create multiple Requests because we currently
210 // don't have the ability to tell if a CompletionProvider::complete response
211 // was a "too many tokens in this request" error. If we had that, then
212 // we could try the request once, instead of having to make separate requests
213 // to check the token count and then afterwards to run the actual prompt.
214 let make_request = |prompt: String| LanguageModelRequest {
215 messages: vec![LanguageModelRequestMessage {
216 role: Role::User,
217 content: vec![prompt.into()],
218 // Nothing in here will benefit from caching
219 cache: false,
220 }],
221 tools: Vec::new(),
222 stop: Vec::new(),
223 temperature: None,
224 };
225
226 while let Some(current_summaries) = stack.pop() {
227 // The split can result in one slice being empty and the other having one element.
228 // Whenever that happens, skip the empty one.
229 if current_summaries.is_empty() {
230 continue;
231 }
232
233 log::info!(
234 "Inferring prompt context using {} file summaries",
235 current_summaries.len()
236 );
237
238 let prompt = summaries_prompt(¤t_summaries, original_prompt);
239 let start = std::time::Instant::now();
240 // Per OpenAI, 1 token ~= 4 chars in English (we go with 4.5 to overestimate a bit, because failed API requests cost a lot of perf)
241 // Verifying this against an actual model.count_tokens() confirms that it's usually within ~5% of the correct answer, whereas
242 // getting the correct answer from tiktoken takes hundreds of milliseconds (compared to this arithmetic being ~free).
243 // source: https://help.openai.com/en/articles/4936856-what-are-tokens-and-how-to-count-them
244 let token_estimate = prompt.len() * 2 / 9;
245 let duration = start.elapsed();
246 log::info!(
247 "Time taken to count tokens for prompt of length {:?}B: {:?}",
248 prompt.len(),
249 duration
250 );
251
252 if token_estimate < max_token_count {
253 prompts.push(prompt);
254 } else if current_summaries.len() == 1 {
255 log::warn!("Inferring context for a single file's summary failed because the prompt's token length exceeded the model's token limit.");
256 } else {
257 log::info!(
258 "Context inference using file summaries resulted in a prompt containing {token_estimate} tokens, which exceeded the model's max of {max_token_count}. Retrying as two separate prompts, each including half the number of summaries.",
259 );
260 let (left, right) = current_summaries.split_at(current_summaries.len() / 2);
261 stack.push(right);
262 stack.push(left);
263 }
264 }
265
266 let all_start = std::time::Instant::now();
267
268 let (tx, rx) = channel::bounded(1024);
269
270 let completion_streams = prompts
271 .into_iter()
272 .map(|prompt| {
273 let request = make_request(prompt.clone());
274 let model = model.clone();
275 let tx = tx.clone();
276 let stream = model.stream_completion(request, &cx);
277
278 (stream, tx)
279 })
280 .collect::<Vec<_>>();
281
282 cx.background_executor()
283 .spawn(async move {
284 let futures = completion_streams
285 .into_iter()
286 .enumerate()
287 .map(|(ix, (stream, tx))| async move {
288 let start = std::time::Instant::now();
289 let events = stream.await?;
290 log::info!("Time taken for awaiting /await chunk stream #{ix}: {:?}", start.elapsed());
291
292 let completion: String = events
293 .filter_map(|event| async {
294 if let Ok(LanguageModelCompletionEvent::Text(text)) = event {
295 Some(text)
296 } else {
297 None
298 }
299 })
300 .collect()
301 .await;
302
303 log::info!("Time taken for all /auto chunks to come back for #{ix}: {:?}", start.elapsed());
304
305 for line in completion.split('\n') {
306 if let Some(first_space) = line.find(' ') {
307 let command = &line[..first_space].trim();
308 let arg = &line[first_space..].trim();
309
310 tx.send(CommandToRun {
311 name: command.to_string(),
312 arg: arg.to_string(),
313 })
314 .await?;
315 } else if !line.trim().is_empty() {
316 // All slash-commands currently supported in context inference need a space for the argument.
317 log::warn!(
318 "Context inference returned a non-blank line that contained no spaces (meaning no argument for the slash command): {:?}",
319 line
320 );
321 }
322 }
323
324 anyhow::Ok(())
325 })
326 .collect::<Vec<_>>();
327
328 let _ = futures::future::try_join_all(futures).await.log_err();
329
330 let duration = all_start.elapsed();
331 eprintln!("All futures completed in {:?}", duration);
332 })
333 .await;
334
335 drop(tx); // Close the channel so that rx.collect() won't hang. This is safe because all futures have completed.
336 let results = rx.collect::<Vec<_>>().await;
337 eprintln!(
338 "Finished collecting from the channel with {} results",
339 results.len()
340 );
341 for command in results {
342 // Don't return empty or duplicate commands
343 if !command.name.is_empty()
344 && !final_response
345 .iter()
346 .any(|cmd: &CommandToRun| cmd.name == command.name && cmd.arg == command.arg)
347 {
348 if SUPPORTED_SLASH_COMMANDS
349 .iter()
350 .any(|supported| &command.name == supported)
351 {
352 final_response.push(command);
353 } else {
354 log::warn!(
355 "Context inference returned an unrecognized slash command: {:?}",
356 command
357 );
358 }
359 }
360 }
361
362 // Sort the commands by name (reversed just so that /search appears before /file)
363 final_response.sort_by(|cmd1, cmd2| cmd1.name.cmp(&cmd2.name).reverse());
364
365 Ok(final_response)
366}