/auto (#16696)
Add `/auto` behind a feature flag that's disabled for now, even for staff. We've decided on a different design for context inference, but there are parts of /auto that will be useful for that, so we want them in the code base even if they're unused for now. Release Notes: - N/A --------- Co-authored-by: Antonio Scandurra <me@as-cii.com> Co-authored-by: Marshall Bowers <elliott.codes@gmail.com>
This commit is contained in:
Richard Feldman
GitHub
parent
93a3e8bc94
commit
91ffa02e2c
42 changed files with 2776 additions and 1054 deletions
360
crates/assistant/src/slash_command/auto_command.rs
Normal file
360
crates/assistant/src/slash_command/auto_command.rs
Normal file
|
|
@ -0,0 +1,360 @@
|
|||
use super::create_label_for_command;
|
||||
use super::{SlashCommand, SlashCommandOutput};
|
||||
use anyhow::{anyhow, Result};
|
||||
use assistant_slash_command::ArgumentCompletion;
|
||||
use feature_flags::FeatureFlag;
|
||||
use futures::StreamExt;
|
||||
use gpui::{AppContext, AsyncAppContext, Task, WeakView};
|
||||
use language::{CodeLabel, LspAdapterDelegate};
|
||||
use language_model::{
|
||||
LanguageModelCompletionEvent, LanguageModelRegistry, LanguageModelRequest,
|
||||
LanguageModelRequestMessage, Role,
|
||||
};
|
||||
use semantic_index::{FileSummary, SemanticDb};
|
||||
use smol::channel;
|
||||
use std::sync::{atomic::AtomicBool, Arc};
|
||||
use ui::{BorrowAppContext, WindowContext};
|
||||
use util::ResultExt;
|
||||
use workspace::Workspace;
|
||||
|
||||
pub struct AutoSlashCommandFeatureFlag;
|
||||
|
||||
impl FeatureFlag for AutoSlashCommandFeatureFlag {
|
||||
const NAME: &'static str = "auto-slash-command";
|
||||
}
|
||||
|
||||
pub(crate) struct AutoCommand;
|
||||
|
||||
impl SlashCommand for AutoCommand {
|
||||
fn name(&self) -> String {
|
||||
"auto".into()
|
||||
}
|
||||
|
||||
fn description(&self) -> String {
|
||||
"Automatically infer what context to add, based on your prompt".into()
|
||||
}
|
||||
|
||||
fn menu_text(&self) -> String {
|
||||
"Automatically Infer Context".into()
|
||||
}
|
||||
|
||||
fn label(&self, cx: &AppContext) -> CodeLabel {
|
||||
create_label_for_command("auto", &["--prompt"], cx)
|
||||
}
|
||||
|
||||
fn complete_argument(
|
||||
self: Arc<Self>,
|
||||
_arguments: &[String],
|
||||
_cancel: Arc<AtomicBool>,
|
||||
workspace: Option<WeakView<Workspace>>,
|
||||
cx: &mut WindowContext,
|
||||
) -> Task<Result<Vec<ArgumentCompletion>>> {
|
||||
// There's no autocomplete for a prompt, since it's arbitrary text.
|
||||
// However, we can use this opportunity to kick off a drain of the backlog.
|
||||
// That way, it can hopefully be done resummarizing by the time we've actually
|
||||
// typed out our prompt. This re-runs on every keystroke during autocomplete,
|
||||
// but in the future, we could instead do it only once, when /auto is first entered.
|
||||
let Some(workspace) = workspace.and_then(|ws| ws.upgrade()) else {
|
||||
log::warn!("workspace was dropped or unavailable during /auto autocomplete");
|
||||
|
||||
return Task::ready(Ok(Vec::new()));
|
||||
};
|
||||
|
||||
let project = workspace.read(cx).project().clone();
|
||||
let Some(project_index) =
|
||||
cx.update_global(|index: &mut SemanticDb, cx| index.project_index(project, cx))
|
||||
else {
|
||||
return Task::ready(Err(anyhow!("No project indexer, cannot use /auto")));
|
||||
};
|
||||
|
||||
let cx: &mut AppContext = cx;
|
||||
|
||||
cx.spawn(|cx: gpui::AsyncAppContext| async move {
|
||||
let task = project_index.read_with(&cx, |project_index, cx| {
|
||||
project_index.flush_summary_backlogs(cx)
|
||||
})?;
|
||||
|
||||
cx.background_executor().spawn(task).await;
|
||||
|
||||
anyhow::Ok(Vec::new())
|
||||
})
|
||||
}
|
||||
|
||||
fn requires_argument(&self) -> bool {
|
||||
true
|
||||
}
|
||||
|
||||
fn run(
|
||||
self: Arc<Self>,
|
||||
arguments: &[String],
|
||||
workspace: WeakView<Workspace>,
|
||||
_delegate: Option<Arc<dyn LspAdapterDelegate>>,
|
||||
cx: &mut WindowContext,
|
||||
) -> Task<Result<SlashCommandOutput>> {
|
||||
let Some(workspace) = workspace.upgrade() else {
|
||||
return Task::ready(Err(anyhow::anyhow!("workspace was dropped")));
|
||||
};
|
||||
if arguments.is_empty() {
|
||||
return Task::ready(Err(anyhow!("missing prompt")));
|
||||
};
|
||||
let argument = arguments.join(" ");
|
||||
let original_prompt = argument.to_string();
|
||||
let project = workspace.read(cx).project().clone();
|
||||
let Some(project_index) =
|
||||
cx.update_global(|index: &mut SemanticDb, cx| index.project_index(project, cx))
|
||||
else {
|
||||
return Task::ready(Err(anyhow!("no project indexer")));
|
||||
};
|
||||
|
||||
let task = cx.spawn(|cx: gpui::AsyncWindowContext| async move {
|
||||
let summaries = project_index
|
||||
.read_with(&cx, |project_index, cx| project_index.all_summaries(cx))?
|
||||
.await?;
|
||||
|
||||
commands_for_summaries(&summaries, &original_prompt, &cx).await
|
||||
});
|
||||
|
||||
// As a convenience, append /auto's argument to the end of the prompt
|
||||
// so you don't have to write it again.
|
||||
let original_prompt = argument.to_string();
|
||||
|
||||
cx.background_executor().spawn(async move {
|
||||
let commands = task.await?;
|
||||
let mut prompt = String::new();
|
||||
|
||||
log::info!(
|
||||
"Translating this response into slash-commands: {:?}",
|
||||
commands
|
||||
);
|
||||
|
||||
for command in commands {
|
||||
prompt.push('/');
|
||||
prompt.push_str(&command.name);
|
||||
prompt.push(' ');
|
||||
prompt.push_str(&command.arg);
|
||||
prompt.push('\n');
|
||||
}
|
||||
|
||||
prompt.push('\n');
|
||||
prompt.push_str(&original_prompt);
|
||||
|
||||
Ok(SlashCommandOutput {
|
||||
text: prompt,
|
||||
sections: Vec::new(),
|
||||
run_commands_in_text: true,
|
||||
})
|
||||
})
|
||||
}
|
||||
}
|
||||
|
||||
const PROMPT_INSTRUCTIONS_BEFORE_SUMMARY: &str = include_str!("prompt_before_summary.txt");
|
||||
const PROMPT_INSTRUCTIONS_AFTER_SUMMARY: &str = include_str!("prompt_after_summary.txt");
|
||||
|
||||
fn summaries_prompt(summaries: &[FileSummary], original_prompt: &str) -> String {
|
||||
let json_summaries = serde_json::to_string(summaries).unwrap();
|
||||
|
||||
format!("{PROMPT_INSTRUCTIONS_BEFORE_SUMMARY}\n{json_summaries}\n{PROMPT_INSTRUCTIONS_AFTER_SUMMARY}\n{original_prompt}")
|
||||
}
|
||||
|
||||
/// The slash commands that the model is told about, and which we look for in the inference response.
|
||||
const SUPPORTED_SLASH_COMMANDS: &[&str] = &["search", "file"];
|
||||
|
||||
#[derive(Debug, Clone)]
|
||||
struct CommandToRun {
|
||||
name: String,
|
||||
arg: String,
|
||||
}
|
||||
|
||||
/// Given the pre-indexed file summaries for this project, as well as the original prompt
|
||||
/// string passed to `/auto`, get a list of slash commands to run, along with their arguments.
|
||||
///
|
||||
/// The prompt's output does not include the slashes (to reduce the chance that it makes a mistake),
|
||||
/// so taking one of these returned Strings and turning it into a real slash-command-with-argument
|
||||
/// involves prepending a slash to it.
|
||||
///
|
||||
/// This function will validate that each of the returned lines begins with one of SUPPORTED_SLASH_COMMANDS.
|
||||
/// Any other lines it encounters will be discarded, with a warning logged.
|
||||
async fn commands_for_summaries(
|
||||
summaries: &[FileSummary],
|
||||
original_prompt: &str,
|
||||
cx: &AsyncAppContext,
|
||||
) -> Result<Vec<CommandToRun>> {
|
||||
if summaries.is_empty() {
|
||||
log::warn!("Inferring no context because there were no summaries available.");
|
||||
return Ok(Vec::new());
|
||||
}
|
||||
|
||||
// Use the globally configured model to translate the summaries into slash-commands,
|
||||
// because Qwen2-7B-Instruct has not done a good job at that task.
|
||||
let Some(model) = cx.update(|cx| LanguageModelRegistry::read_global(cx).active_model())? else {
|
||||
log::warn!("Can't infer context because there's no active model.");
|
||||
return Ok(Vec::new());
|
||||
};
|
||||
// Only go up to 90% of the actual max token count, to reduce chances of
|
||||
// exceeding the token count due to inaccuracies in the token counting heuristic.
|
||||
let max_token_count = (model.max_token_count() * 9) / 10;
|
||||
|
||||
// Rather than recursing (which would require this async function use a pinned box),
|
||||
// we use an explicit stack of arguments and answers for when we need to "recurse."
|
||||
let mut stack = vec![summaries];
|
||||
let mut final_response = Vec::new();
|
||||
let mut prompts = Vec::new();
|
||||
|
||||
// TODO We only need to create multiple Requests because we currently
|
||||
// don't have the ability to tell if a CompletionProvider::complete response
|
||||
// was a "too many tokens in this request" error. If we had that, then
|
||||
// we could try the request once, instead of having to make separate requests
|
||||
// to check the token count and then afterwards to run the actual prompt.
|
||||
let make_request = |prompt: String| LanguageModelRequest {
|
||||
messages: vec![LanguageModelRequestMessage {
|
||||
role: Role::User,
|
||||
content: vec![prompt.into()],
|
||||
// Nothing in here will benefit from caching
|
||||
cache: false,
|
||||
}],
|
||||
tools: Vec::new(),
|
||||
stop: Vec::new(),
|
||||
temperature: 1.0,
|
||||
};
|
||||
|
||||
while let Some(current_summaries) = stack.pop() {
|
||||
// The split can result in one slice being empty and the other having one element.
|
||||
// Whenever that happens, skip the empty one.
|
||||
if current_summaries.is_empty() {
|
||||
continue;
|
||||
}
|
||||
|
||||
log::info!(
|
||||
"Inferring prompt context using {} file summaries",
|
||||
current_summaries.len()
|
||||
);
|
||||
|
||||
let prompt = summaries_prompt(¤t_summaries, original_prompt);
|
||||
let start = std::time::Instant::now();
|
||||
// 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)
|
||||
// Verifying this against an actual model.count_tokens() confirms that it's usually within ~5% of the correct answer, whereas
|
||||
// getting the correct answer from tiktoken takes hundreds of milliseconds (compared to this arithmetic being ~free).
|
||||
// source: https://help.openai.com/en/articles/4936856-what-are-tokens-and-how-to-count-them
|
||||
let token_estimate = prompt.len() * 2 / 9;
|
||||
let duration = start.elapsed();
|
||||
log::info!(
|
||||
"Time taken to count tokens for prompt of length {:?}B: {:?}",
|
||||
prompt.len(),
|
||||
duration
|
||||
);
|
||||
|
||||
if token_estimate < max_token_count {
|
||||
prompts.push(prompt);
|
||||
} else if current_summaries.len() == 1 {
|
||||
log::warn!("Inferring context for a single file's summary failed because the prompt's token length exceeded the model's token limit.");
|
||||
} else {
|
||||
log::info!(
|
||||
"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.",
|
||||
);
|
||||
let (left, right) = current_summaries.split_at(current_summaries.len() / 2);
|
||||
stack.push(right);
|
||||
stack.push(left);
|
||||
}
|
||||
}
|
||||
|
||||
let all_start = std::time::Instant::now();
|
||||
|
||||
let (tx, rx) = channel::bounded(1024);
|
||||
|
||||
let completion_streams = prompts
|
||||
.into_iter()
|
||||
.map(|prompt| {
|
||||
let request = make_request(prompt.clone());
|
||||
let model = model.clone();
|
||||
let tx = tx.clone();
|
||||
let stream = model.stream_completion(request, &cx);
|
||||
|
||||
(stream, tx)
|
||||
})
|
||||
.collect::<Vec<_>>();
|
||||
|
||||
cx.background_executor()
|
||||
.spawn(async move {
|
||||
let futures = completion_streams
|
||||
.into_iter()
|
||||
.enumerate()
|
||||
.map(|(ix, (stream, tx))| async move {
|
||||
let start = std::time::Instant::now();
|
||||
let events = stream.await?;
|
||||
log::info!("Time taken for awaiting /await chunk stream #{ix}: {:?}", start.elapsed());
|
||||
|
||||
let completion: String = events
|
||||
.filter_map(|event| async {
|
||||
if let Ok(LanguageModelCompletionEvent::Text(text)) = event {
|
||||
Some(text)
|
||||
} else {
|
||||
None
|
||||
}
|
||||
})
|
||||
.collect()
|
||||
.await;
|
||||
|
||||
log::info!("Time taken for all /auto chunks to come back for #{ix}: {:?}", start.elapsed());
|
||||
|
||||
for line in completion.split('\n') {
|
||||
if let Some(first_space) = line.find(' ') {
|
||||
let command = &line[..first_space].trim();
|
||||
let arg = &line[first_space..].trim();
|
||||
|
||||
tx.send(CommandToRun {
|
||||
name: command.to_string(),
|
||||
arg: arg.to_string(),
|
||||
})
|
||||
.await?;
|
||||
} else if !line.trim().is_empty() {
|
||||
// All slash-commands currently supported in context inference need a space for the argument.
|
||||
log::warn!(
|
||||
"Context inference returned a non-blank line that contained no spaces (meaning no argument for the slash command): {:?}",
|
||||
line
|
||||
);
|
||||
}
|
||||
}
|
||||
|
||||
anyhow::Ok(())
|
||||
})
|
||||
.collect::<Vec<_>>();
|
||||
|
||||
let _ = futures::future::try_join_all(futures).await.log_err();
|
||||
|
||||
let duration = all_start.elapsed();
|
||||
eprintln!("All futures completed in {:?}", duration);
|
||||
})
|
||||
.await;
|
||||
|
||||
drop(tx); // Close the channel so that rx.collect() won't hang. This is safe because all futures have completed.
|
||||
let results = rx.collect::<Vec<_>>().await;
|
||||
eprintln!(
|
||||
"Finished collecting from the channel with {} results",
|
||||
results.len()
|
||||
);
|
||||
for command in results {
|
||||
// Don't return empty or duplicate commands
|
||||
if !command.name.is_empty()
|
||||
&& !final_response
|
||||
.iter()
|
||||
.any(|cmd: &CommandToRun| cmd.name == command.name && cmd.arg == command.arg)
|
||||
{
|
||||
if SUPPORTED_SLASH_COMMANDS
|
||||
.iter()
|
||||
.any(|supported| &command.name == supported)
|
||||
{
|
||||
final_response.push(command);
|
||||
} else {
|
||||
log::warn!(
|
||||
"Context inference returned an unrecognized slash command: {:?}",
|
||||
command
|
||||
);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Sort the commands by name (reversed just so that /search appears before /file)
|
||||
final_response.sort_by(|cmd1, cmd2| cmd1.name.cmp(&cmd2.name).reverse());
|
||||
|
||||
Ok(final_response)
|
||||
}
|
||||
Loading…
Add table
Add a link
Reference in a new issue