Allow AI interactions to be proxied through Zed's server so you don't need an API key (#7367)

Co-authored-by: Antonio <antonio@zed.dev>

Resurrected this from some assistant work I did in Spring of 2023.
- [x] Resurrect streaming responses
- [x] Use streaming responses to enable AI via Zed's servers by default
(but preserve API key option for now)
- [x] Simplify protobuf
- [x] Proxy to OpenAI on zed.dev
- [x] Proxy to Gemini on zed.dev
- [x] Improve UX for switching between openAI and google models
- We current disallow cycling when setting a custom model, but we need a
better solution to keep OpenAI models available while testing the google
ones
- [x] Show remaining tokens correctly for Google models
- [x] Remove semantic index
- [x] Delete `ai` crate
- [x] Cloud front so we can ban abuse
- [x] Rate-limiting
- [x] Fix panic when using inline assistant
- [x] Double check the upgraded `AssistantSettings` are
backwards-compatible
- [x] Add hosted LLM interaction behind a `language-models` feature
flag.

Release Notes:

- We are temporarily removing the semantic index in order to redesign it
from scratch.

---------

Co-authored-by: Antonio <antonio@zed.dev>
Co-authored-by: Antonio Scandurra <me@as-cii.com>
Co-authored-by: Thorsten <thorsten@zed.dev>
Co-authored-by: Max <max@zed.dev>
This commit is contained in:
Nathan Sobo 2024-03-19 12:22:26 -06:00 committed by GitHub
parent 905a24079a
commit 8ae5a3b61a
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
87 changed files with 3647 additions and 8937 deletions

View file

@ -1,394 +1,95 @@
use ai::models::LanguageModel;
use ai::prompts::base::{PromptArguments, PromptChain, PromptPriority, PromptTemplate};
use ai::prompts::file_context::FileContext;
use ai::prompts::generate::GenerateInlineContent;
use ai::prompts::preamble::EngineerPreamble;
use ai::prompts::repository_context::{PromptCodeSnippet, RepositoryContext};
use ai::providers::open_ai::OpenAiLanguageModel;
use language::{BufferSnapshot, OffsetRangeExt, ToOffset};
use std::cmp::{self, Reverse};
use std::ops::Range;
use std::sync::Arc;
#[allow(dead_code)]
fn summarize(buffer: &BufferSnapshot, selected_range: Range<impl ToOffset>) -> String {
#[derive(Debug)]
struct Match {
collapse: Range<usize>,
keep: Vec<Range<usize>>,
}
let selected_range = selected_range.to_offset(buffer);
let mut ts_matches = buffer.matches(0..buffer.len(), |grammar| {
Some(&grammar.embedding_config.as_ref()?.query)
});
let configs = ts_matches
.grammars()
.iter()
.map(|g| g.embedding_config.as_ref().unwrap())
.collect::<Vec<_>>();
let mut matches = Vec::new();
while let Some(mat) = ts_matches.peek() {
let config = &configs[mat.grammar_index];
if let Some(collapse) = mat.captures.iter().find_map(|cap| {
if Some(cap.index) == config.collapse_capture_ix {
Some(cap.node.byte_range())
} else {
None
}
}) {
let mut keep = Vec::new();
for capture in mat.captures.iter() {
if Some(capture.index) == config.keep_capture_ix {
keep.push(capture.node.byte_range());
} else {
continue;
}
}
ts_matches.advance();
matches.push(Match { collapse, keep });
} else {
ts_matches.advance();
}
}
matches.sort_unstable_by_key(|mat| (mat.collapse.start, Reverse(mat.collapse.end)));
let mut matches = matches.into_iter().peekable();
let mut summary = String::new();
let mut offset = 0;
let mut flushed_selection = false;
while let Some(mat) = matches.next() {
// Keep extending the collapsed range if the next match surrounds
// the current one.
while let Some(next_mat) = matches.peek() {
if mat.collapse.start <= next_mat.collapse.start
&& mat.collapse.end >= next_mat.collapse.end
{
matches.next().unwrap();
} else {
break;
}
}
if offset > mat.collapse.start {
// Skip collapsed nodes that have already been summarized.
offset = cmp::max(offset, mat.collapse.end);
continue;
}
if offset <= selected_range.start && selected_range.start <= mat.collapse.end {
if !flushed_selection {
// The collapsed node ends after the selection starts, so we'll flush the selection first.
summary.extend(buffer.text_for_range(offset..selected_range.start));
summary.push_str("<|S|");
if selected_range.end == selected_range.start {
summary.push_str(">");
} else {
summary.extend(buffer.text_for_range(selected_range.clone()));
summary.push_str("|E|>");
}
offset = selected_range.end;
flushed_selection = true;
}
// If the selection intersects the collapsed node, we won't collapse it.
if selected_range.end >= mat.collapse.start {
continue;
}
}
summary.extend(buffer.text_for_range(offset..mat.collapse.start));
for keep in mat.keep {
summary.extend(buffer.text_for_range(keep));
}
offset = mat.collapse.end;
}
// Flush selection if we haven't already done so.
if !flushed_selection && offset <= selected_range.start {
summary.extend(buffer.text_for_range(offset..selected_range.start));
summary.push_str("<|S|");
if selected_range.end == selected_range.start {
summary.push_str(">");
} else {
summary.extend(buffer.text_for_range(selected_range.clone()));
summary.push_str("|E|>");
}
offset = selected_range.end;
}
summary.extend(buffer.text_for_range(offset..buffer.len()));
summary
}
use language::BufferSnapshot;
use std::{fmt::Write, ops::Range};
pub fn generate_content_prompt(
user_prompt: String,
language_name: Option<&str>,
buffer: BufferSnapshot,
range: Range<usize>,
search_results: Vec<PromptCodeSnippet>,
model: &str,
project_name: Option<String>,
) -> anyhow::Result<String> {
// Using new Prompt Templates
let openai_model: Arc<dyn LanguageModel> = Arc::new(OpenAiLanguageModel::load(model));
let lang_name = if let Some(language_name) = language_name {
Some(language_name.to_string())
let mut prompt = String::new();
let content_type = match language_name {
None | Some("Markdown" | "Plain Text") => {
writeln!(prompt, "You are an expert engineer.")?;
"Text"
}
Some(language_name) => {
writeln!(prompt, "You are an expert {language_name} engineer.")?;
writeln!(
prompt,
"Your answer MUST always and only be valid {}.",
language_name
)?;
"Code"
}
};
if let Some(project_name) = project_name {
writeln!(
prompt,
"You are currently working inside the '{project_name}' project in code editor Zed."
)?;
}
// Include file content.
for chunk in buffer.text_for_range(0..range.start) {
prompt.push_str(chunk);
}
if range.is_empty() {
prompt.push_str("<|START|>");
} else {
None
};
let args = PromptArguments {
model: openai_model,
language_name: lang_name.clone(),
project_name,
snippets: search_results.clone(),
reserved_tokens: 1000,
buffer: Some(buffer),
selected_range: Some(range),
user_prompt: Some(user_prompt.clone()),
};
let templates: Vec<(PromptPriority, Box<dyn PromptTemplate>)> = vec![
(PromptPriority::Mandatory, Box::new(EngineerPreamble {})),
(
PromptPriority::Ordered { order: 1 },
Box::new(RepositoryContext {}),
),
(
PromptPriority::Ordered { order: 0 },
Box::new(FileContext {}),
),
(
PromptPriority::Mandatory,
Box::new(GenerateInlineContent {}),
),
];
let chain = PromptChain::new(args, templates);
let (prompt, _) = chain.generate(true)?;
anyhow::Ok(prompt)
}
#[cfg(test)]
pub(crate) mod tests {
use super::*;
use gpui::{AppContext, Context};
use indoc::indoc;
use language::{
language_settings, tree_sitter_rust, Buffer, BufferId, Language, LanguageConfig,
LanguageMatcher, Point,
};
use settings::SettingsStore;
use std::sync::Arc;
pub(crate) fn rust_lang() -> Language {
Language::new(
LanguageConfig {
name: "Rust".into(),
matcher: LanguageMatcher {
path_suffixes: vec!["rs".to_string()],
..Default::default()
},
..Default::default()
},
Some(tree_sitter_rust::language()),
)
.with_embedding_query(
r#"
(
[(line_comment) (attribute_item)]* @context
.
[
(struct_item
name: (_) @name)
(enum_item
name: (_) @name)
(impl_item
trait: (_)? @name
"for"? @name
type: (_) @name)
(trait_item
name: (_) @name)
(function_item
name: (_) @name
body: (block
"{" @keep
"}" @keep) @collapse)
(macro_definition
name: (_) @name)
] @item
)
"#,
)
.unwrap()
prompt.push_str("<|START|");
}
#[gpui::test]
fn test_outline_for_prompt(cx: &mut AppContext) {
let settings_store = SettingsStore::test(cx);
cx.set_global(settings_store);
language_settings::init(cx);
let text = indoc! {"
struct X {
a: usize,
b: usize,
}
impl X {
fn new() -> Self {
let a = 1;
let b = 2;
Self { a, b }
}
pub fn a(&self, param: bool) -> usize {
self.a
}
pub fn b(&self) -> usize {
self.b
}
}
"};
let buffer = cx.new_model(|cx| {
Buffer::new(0, BufferId::new(1).unwrap(), text).with_language(Arc::new(rust_lang()), cx)
});
let snapshot = buffer.read(cx).snapshot();
assert_eq!(
summarize(&snapshot, Point::new(1, 4)..Point::new(1, 4)),
indoc! {"
struct X {
<|S|>a: usize,
b: usize,
}
impl X {
fn new() -> Self {}
pub fn a(&self, param: bool) -> usize {}
pub fn b(&self) -> usize {}
}
"}
);
assert_eq!(
summarize(&snapshot, Point::new(8, 12)..Point::new(8, 14)),
indoc! {"
struct X {
a: usize,
b: usize,
}
impl X {
fn new() -> Self {
let <|S|a |E|>= 1;
let b = 2;
Self { a, b }
}
pub fn a(&self, param: bool) -> usize {}
pub fn b(&self) -> usize {}
}
"}
);
assert_eq!(
summarize(&snapshot, Point::new(6, 0)..Point::new(6, 0)),
indoc! {"
struct X {
a: usize,
b: usize,
}
impl X {
<|S|>
fn new() -> Self {}
pub fn a(&self, param: bool) -> usize {}
pub fn b(&self) -> usize {}
}
"}
);
assert_eq!(
summarize(&snapshot, Point::new(21, 0)..Point::new(21, 0)),
indoc! {"
struct X {
a: usize,
b: usize,
}
impl X {
fn new() -> Self {}
pub fn a(&self, param: bool) -> usize {}
pub fn b(&self) -> usize {}
}
<|S|>"}
);
// Ensure nested functions get collapsed properly.
let text = indoc! {"
struct X {
a: usize,
b: usize,
}
impl X {
fn new() -> Self {
let a = 1;
let b = 2;
Self { a, b }
}
pub fn a(&self, param: bool) -> usize {
let a = 30;
fn nested() -> usize {
3
}
self.a + nested()
}
pub fn b(&self) -> usize {
self.b
}
}
"};
buffer.update(cx, |buffer, cx| buffer.set_text(text, cx));
let snapshot = buffer.read(cx).snapshot();
assert_eq!(
summarize(&snapshot, Point::new(0, 0)..Point::new(0, 0)),
indoc! {"
<|S|>struct X {
a: usize,
b: usize,
}
impl X {
fn new() -> Self {}
pub fn a(&self, param: bool) -> usize {}
pub fn b(&self) -> usize {}
}
"}
);
for chunk in buffer.text_for_range(range.clone()) {
prompt.push_str(chunk);
}
if !range.is_empty() {
prompt.push_str("|END|>");
}
for chunk in buffer.text_for_range(range.end..buffer.len()) {
prompt.push_str(chunk);
}
prompt.push('\n');
if range.is_empty() {
writeln!(
prompt,
"Assume the cursor is located where the `<|START|>` span is."
)
.unwrap();
writeln!(
prompt,
"{content_type} can't be replaced, so assume your answer will be inserted at the cursor.",
)
.unwrap();
writeln!(
prompt,
"Generate {content_type} based on the users prompt: {user_prompt}",
)
.unwrap();
} else {
writeln!(prompt, "Modify the user's selected {content_type} based upon the users prompt: '{user_prompt}'").unwrap();
writeln!(prompt, "You must reply with only the adjusted {content_type} (within the '<|START|' and '|END|>' spans) not the entire file.").unwrap();
writeln!(
prompt,
"Double check that you only return code and not the '<|START|' and '|END|'> spans"
)
.unwrap();
}
writeln!(prompt, "Never make remarks about the output.").unwrap();
writeln!(
prompt,
"Do not return anything else, except the generated {content_type}."
)
.unwrap();
Ok(prompt)
}