zed-industries/zed
1
1
Fork 0
mirror of https://github.com/zed-industries/zed.git synced 2024-09-18 18:08:07 +03:00
Code Issues Packages Projects Releases Wiki Activity

Improve prompt for Claude models (#13531)

Browse Source 
This inline assistant prompt is one I designed that in my experience
works much better with Claude 3.5 Sonnet than the default prompt.

Mainly because it takes advantage of a weird property of our finetuning
which is that when you use XML tags it knows that it's doing a
machine-read tasks and stops trying to elide things for brevity. The
default prompt will often remove comments and otherwise add elisions for
brevity when doing large rewrites.

It also avoids giving the entire file content twice when the rewrite
region is large relative to the non-rewritten region.

Not necessarily meant to be merged as-is since it may mess up OAI
models. This is mainly meant for your reference. But everyone should be
using 3.5 Sonnet for coding use cases now anyhow 😛

Release Notes:

- N/A
This commit is contained in:
Tristan Hume 2024-06-26 14:41:40 -04:00 committed by GitHub
parent 4d5441c09d
commit c0a3642f77
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 82 additions and 94 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

176
crates/assistant/src/prompts.rs
View File

@ -6,118 +6,106 @@ pub fn generate_content_prompt(
language_name: Option<&str>,
buffer: BufferSnapshot,
range: Range<usize>,
project_name: Option<String>,
_project_name: Option<String>,
) -> anyhow::Result<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
"Here's a file of text that I'm going to ask you to make an edit to."
)?;
"Code"
"text"
}
Some(language_name) => {
writeln!(
prompt,
"Here's a file of {language_name} that I'm going to ask you to make an edit to."
)?;
"code"
}
};
if let Some(project_name) = project_name {
writeln!(
prompt,
"You are currently working inside the '{project_name}' project in code editor Zed."
)?;
}
writeln!(
prompt,
"The user has the following file open in the editor:"
)?;
const MAX_CTX: usize = 50000;
let mut is_truncated = false;
if range.is_empty() {
write!(prompt, "```")?;
if let Some(language_name) = language_name {
write!(prompt, "{language_name}")?;
}
for chunk in buffer.as_rope().chunks_in_range(0..range.start) {
prompt.push_str(chunk);
}
prompt.push_str("<|CURSOR|>");
for chunk in buffer.as_rope().chunks_in_range(range.start..buffer.len()) {
prompt.push_str(chunk);
}
if !prompt.ends_with('\n') {
prompt.push('\n');
}
writeln!(prompt, "```")?;
prompt.push('\n');
writeln!(
prompt,
"Assume the cursor is located where the `<|CURSOR|>` 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();
prompt.push_str("The point you'll need to insert at is marked with <insert_here></insert_here>.\n\n<document>");
} else {
write!(prompt, "```")?;
for chunk in buffer.as_rope().chunks() {
prompt.push_str(chunk);
}
if !prompt.ends_with('\n') {
prompt.push('\n');
}
writeln!(prompt, "```")?;
prompt.push('\n');
writeln!(
prompt,
"In particular, the following piece of text is selected:"
)?;
write!(prompt, "```")?;
if let Some(language_name) = language_name {
write!(prompt, "{language_name}")?;
}
prompt.push('\n');
prompt.push_str("The section you'll need to rewrite is marked with <rewrite_this></rewrite_this> tags.\n\n<document>");
}
// Include file content.
let before_range = 0..range.start;
let truncated_before = if before_range.len() > MAX_CTX {
is_truncated = true;
range.start - MAX_CTX..range.start
} else {
before_range
};
let mut non_rewrite_len = truncated_before.len();
for chunk in buffer.text_for_range(truncated_before) {
prompt.push_str(chunk);
}
if !range.is_empty() {
prompt.push_str("<rewrite_this>\n");
for chunk in buffer.text_for_range(range.clone()) {
prompt.push_str(chunk);
}
if !prompt.ends_with('\n') {
prompt.push('\n');
}
writeln!(prompt, "```")?;
prompt.push('\n');
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}, not the entire file."
)
.unwrap();
prompt.push_str("\n<rewrite_this>");
} else {
prompt.push_str("<insert_here></insert_here>");
}
let after_range = range.end..buffer.len();
let truncated_after = if after_range.len() > MAX_CTX {
is_truncated = true;
range.end..range.end + MAX_CTX
} else {
after_range
};
non_rewrite_len += truncated_after.len();
for chunk in buffer.text_for_range(truncated_after) {
prompt.push_str(chunk);
}
writeln!(prompt, "Never make remarks about the output.").unwrap();
writeln!(
prompt,
"Do not return anything else, except the generated {content_type}."
)
.unwrap();
write!(prompt, "</document>\n\n").unwrap();
if is_truncated {
writeln!(prompt, "The context around the relevant section has been truncated (possibly in the middle of a line) for brevity.\n")?;
}
if range.is_empty() {
writeln!(
prompt,
"You can't replace {content_type}, your answer will be inserted in place of the `<insert_here></insert_here>` tags. Don't include the insert_here tags in your output.",
)
.unwrap();
writeln!(
prompt,
"Generate {content_type} based on the following prompt:\n\n<prompt>\n{user_prompt}\n</prompt>",
)
.unwrap();
writeln!(prompt, "Match the indentation in the original file in the inserted {content_type}, don't include any indentation on blank lines.\n").unwrap();
prompt.push_str("Immediately start with the following format with no remarks:\n\n```\n{{INSERTED_CODE}}\n```");
} else {
writeln!(prompt, "Edit the section of {content_type} in <rewrite_this></rewrite_this> tags based on the following prompt:'").unwrap();
writeln!(prompt, "\n<prompt>\n{user_prompt}\n</prompt>\n").unwrap();
let rewrite_len = range.end - range.start;
if rewrite_len < 20000 && rewrite_len * 2 < non_rewrite_len {
writeln!(prompt, "And here's the section to rewrite based on that prompt again for reference:\n\n<rewrite_this>\n").unwrap();
for chunk in buffer.text_for_range(range.clone()) {
prompt.push_str(chunk);
}
writeln!(prompt, "\n</rewrite_this>\n").unwrap();
}
writeln!(prompt, "Only make changes that are necessary to fulfill the prompt, leave everything else as-is. All surrounding {content_type} will be preserved.\n").unwrap();
write!(
prompt,
"Start at the indentation level in the original file in the rewritten {content_type}. "
)
.unwrap();
prompt.push_str("Don't stop until you've rewritten the entire section, even if you have no more changes to make, always write out the whole section with no unnecessary elisions.");
prompt.push_str("\n\nImmediately start with the following format with no remarks:\n\n```\n{{REWRITTEN_CODE}}\n```");
}
Ok(prompt)
}