Files
oak-gpui/crates/open_ai/src/open_ai.rs
T
邻二氮杂菲 f1778dd9de Add max_output_tokens to OpenAI models and integrate into requests (#16381)
### Pull Request Title
Introduce `max_output_tokens` Field for OpenAI Models


https://platform.deepseek.com/api-docs/news/news0725/#4-8k-max_tokens-betarelease-longer-possibilities

### Description
This commit introduces a new field `max_output_tokens` to the OpenAI
models, which allows specifying the maximum number of tokens that can be
generated in the output. This field is now integrated into the request
handling across multiple crates, ensuring that the output token limit is
respected during language model completions.

Changes include:
- Adding `max_output_tokens` to the `Custom` variant of the
`open_ai::Model` enum.
- Updating the `into_open_ai` method in `LanguageModelRequest` to accept
and use `max_output_tokens`.
- Modifying the `OpenAiLanguageModel` and `CloudLanguageModel`
implementations to pass `max_output_tokens` when converting requests.
- Ensuring that the `max_output_tokens` field is correctly serialized
and deserialized in relevant structures.

This enhancement provides more control over the output length of OpenAI
model responses, improving the flexibility and accuracy of language
model interactions.

### Changes
- Added `max_output_tokens` to the `Custom` variant of the
`open_ai::Model` enum.
- Updated the `into_open_ai` method in `LanguageModelRequest` to accept
and use `max_output_tokens`.
- Modified the `OpenAiLanguageModel` and `CloudLanguageModel`
implementations to pass `max_output_tokens` when converting requests.
- Ensured that the `max_output_tokens` field is correctly serialized and
deserialized in relevant structures.

### Related Issue
https://github.com/zed-industries/zed/pull/16358

### Screenshots / Media
N/A

### Checklist
- [x] Code compiles correctly.
- [x] All tests pass.
- [ ] Documentation has been updated accordingly.
- [ ] Additional tests have been added to cover new functionality.
- [ ] Relevant documentation has been updated or added.

### Release Notes

- Added `max_output_tokens` field to OpenAI models for controlling
output token length.
2024-08-21 00:39:10 -04:00

465 lines
14 KiB
Rust

mod supported_countries;
use anyhow::{anyhow, Context, Result};
use futures::{io::BufReader, stream::BoxStream, AsyncBufReadExt, AsyncReadExt, Stream, StreamExt};
use http_client::{AsyncBody, HttpClient, Method, Request as HttpRequest};
use isahc::config::Configurable;
use serde::{Deserialize, Serialize};
use serde_json::Value;
use std::{convert::TryFrom, future::Future, pin::Pin, time::Duration};
use strum::EnumIter;
pub use supported_countries::*;
pub const OPEN_AI_API_URL: &str = "https://api.openai.com/v1";
fn is_none_or_empty<T: AsRef<[U]>, U>(opt: &Option<T>) -> bool {
opt.as_ref().map_or(true, |v| v.as_ref().is_empty())
}
#[derive(Clone, Copy, Serialize, Deserialize, Debug, Eq, PartialEq)]
#[serde(rename_all = "lowercase")]
pub enum Role {
User,
Assistant,
System,
Tool,
}
impl TryFrom<String> for Role {
type Error = anyhow::Error;
fn try_from(value: String) -> Result<Self> {
match value.as_str() {
"user" => Ok(Self::User),
"assistant" => Ok(Self::Assistant),
"system" => Ok(Self::System),
"tool" => Ok(Self::Tool),
_ => Err(anyhow!("invalid role '{value}'")),
}
}
}
impl From<Role> for String {
fn from(val: Role) -> Self {
match val {
Role::User => "user".to_owned(),
Role::Assistant => "assistant".to_owned(),
Role::System => "system".to_owned(),
Role::Tool => "tool".to_owned(),
}
}
}
#[cfg_attr(feature = "schemars", derive(schemars::JsonSchema))]
#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq, EnumIter)]
pub enum Model {
#[serde(rename = "gpt-3.5-turbo", alias = "gpt-3.5-turbo-0613")]
ThreePointFiveTurbo,
#[serde(rename = "gpt-4", alias = "gpt-4-0613")]
Four,
#[serde(rename = "gpt-4-turbo-preview", alias = "gpt-4-1106-preview")]
FourTurbo,
#[serde(rename = "gpt-4o", alias = "gpt-4o-2024-05-13")]
#[default]
FourOmni,
#[serde(rename = "gpt-4o-mini", alias = "gpt-4o-mini-2024-07-18")]
FourOmniMini,
#[serde(rename = "custom")]
Custom {
name: String,
max_tokens: usize,
max_output_tokens: Option<u32>,
},
}
impl Model {
pub fn from_id(id: &str) -> Result<Self> {
match id {
"gpt-3.5-turbo" => Ok(Self::ThreePointFiveTurbo),
"gpt-4" => Ok(Self::Four),
"gpt-4-turbo-preview" => Ok(Self::FourTurbo),
"gpt-4o" => Ok(Self::FourOmni),
"gpt-4o-mini" => Ok(Self::FourOmniMini),
_ => Err(anyhow!("invalid model id")),
}
}
pub fn id(&self) -> &str {
match self {
Self::ThreePointFiveTurbo => "gpt-3.5-turbo",
Self::Four => "gpt-4",
Self::FourTurbo => "gpt-4-turbo-preview",
Self::FourOmni => "gpt-4o",
Self::FourOmniMini => "gpt-4o-mini",
Self::Custom { name, .. } => name,
}
}
pub fn display_name(&self) -> &str {
match self {
Self::ThreePointFiveTurbo => "gpt-3.5-turbo",
Self::Four => "gpt-4",
Self::FourTurbo => "gpt-4-turbo",
Self::FourOmni => "gpt-4o",
Self::FourOmniMini => "gpt-4o-mini",
Self::Custom { name, .. } => name,
}
}
pub fn max_token_count(&self) -> usize {
match self {
Self::ThreePointFiveTurbo => 4096,
Self::Four => 8192,
Self::FourTurbo => 128000,
Self::FourOmni => 128000,
Self::FourOmniMini => 128000,
Self::Custom { max_tokens, .. } => *max_tokens,
}
}
pub fn max_output_tokens(&self) -> Option<u32> {
match self {
Self::ThreePointFiveTurbo => Some(4096),
Self::Four => Some(8192),
Self::FourTurbo => Some(4096),
Self::FourOmni => Some(4096),
Self::FourOmniMini => Some(16384),
Self::Custom {
max_output_tokens, ..
} => *max_output_tokens,
}
}
}
#[derive(Debug, Serialize, Deserialize)]
pub struct Request {
pub model: String,
pub messages: Vec<RequestMessage>,
pub stream: bool,
#[serde(default, skip_serializing_if = "Option::is_none")]
pub max_tokens: Option<u32>,
pub stop: Vec<String>,
pub temperature: f32,
#[serde(default, skip_serializing_if = "Option::is_none")]
pub tool_choice: Option<ToolChoice>,
#[serde(default, skip_serializing_if = "Vec::is_empty")]
pub tools: Vec<ToolDefinition>,
}
#[derive(Debug, Serialize, Deserialize)]
#[serde(untagged)]
pub enum ToolChoice {
Auto,
Required,
None,
Other(ToolDefinition),
}
#[derive(Clone, Deserialize, Serialize, Debug)]
#[serde(tag = "type", rename_all = "snake_case")]
pub enum ToolDefinition {
#[allow(dead_code)]
Function { function: FunctionDefinition },
}
#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct FunctionDefinition {
pub name: String,
pub description: Option<String>,
pub parameters: Option<Value>,
}
#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
#[serde(tag = "role", rename_all = "lowercase")]
pub enum RequestMessage {
Assistant {
content: Option<String>,
#[serde(default, skip_serializing_if = "Vec::is_empty")]
tool_calls: Vec<ToolCall>,
},
User {
content: String,
},
System {
content: String,
},
Tool {
content: String,
tool_call_id: String,
},
}
#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
pub struct ToolCall {
pub id: String,
#[serde(flatten)]
pub content: ToolCallContent,
}
#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
#[serde(tag = "type", rename_all = "lowercase")]
pub enum ToolCallContent {
Function { function: FunctionContent },
}
#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
pub struct FunctionContent {
pub name: String,
pub arguments: String,
}
#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
pub struct ResponseMessageDelta {
pub role: Option<Role>,
pub content: Option<String>,
#[serde(default, skip_serializing_if = "is_none_or_empty")]
pub tool_calls: Option<Vec<ToolCallChunk>>,
}
#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
pub struct ToolCallChunk {
pub index: usize,
pub id: Option<String>,
// There is also an optional `type` field that would determine if a
// function is there. Sometimes this streams in with the `function` before
// it streams in the `type`
pub function: Option<FunctionChunk>,
}
#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
pub struct FunctionChunk {
pub name: Option<String>,
pub arguments: Option<String>,
}
#[derive(Serialize, Deserialize, Debug)]
pub struct Usage {
pub prompt_tokens: u32,
pub completion_tokens: u32,
pub total_tokens: u32,
}
#[derive(Serialize, Deserialize, Debug)]
pub struct ChoiceDelta {
pub index: u32,
pub delta: ResponseMessageDelta,
pub finish_reason: Option<String>,
}
#[derive(Serialize, Deserialize, Debug)]
#[serde(untagged)]
pub enum ResponseStreamResult {
Ok(ResponseStreamEvent),
Err { error: String },
}
#[derive(Serialize, Deserialize, Debug)]
pub struct ResponseStreamEvent {
pub created: u32,
pub model: String,
pub choices: Vec<ChoiceDelta>,
pub usage: Option<Usage>,
}
pub async fn stream_completion(
client: &dyn HttpClient,
api_url: &str,
api_key: &str,
request: Request,
low_speed_timeout: Option<Duration>,
) -> Result<BoxStream<'static, Result<ResponseStreamEvent>>> {
let uri = format!("{api_url}/chat/completions");
let mut request_builder = HttpRequest::builder()
.method(Method::POST)
.uri(uri)
.header("Content-Type", "application/json")
.header("Authorization", format!("Bearer {}", api_key));
if let Some(low_speed_timeout) = low_speed_timeout {
request_builder = request_builder.low_speed_timeout(100, low_speed_timeout);
};
let request = request_builder.body(AsyncBody::from(serde_json::to_string(&request)?))?;
let mut response = client.send(request).await?;
if response.status().is_success() {
let reader = BufReader::new(response.into_body());
Ok(reader
.lines()
.filter_map(|line| async move {
match line {
Ok(line) => {
let line = line.strip_prefix("data: ")?;
if line == "[DONE]" {
None
} else {
match serde_json::from_str(line) {
Ok(ResponseStreamResult::Ok(response)) => Some(Ok(response)),
Ok(ResponseStreamResult::Err { error }) => {
Some(Err(anyhow!(error)))
}
Err(error) => Some(Err(anyhow!(error))),
}
}
}
Err(error) => Some(Err(anyhow!(error))),
}
})
.boxed())
} else {
let mut body = String::new();
response.body_mut().read_to_string(&mut body).await?;
#[derive(Deserialize)]
struct OpenAiResponse {
error: OpenAiError,
}
#[derive(Deserialize)]
struct OpenAiError {
message: String,
}
match serde_json::from_str::<OpenAiResponse>(&body) {
Ok(response) if !response.error.message.is_empty() => Err(anyhow!(
"Failed to connect to OpenAI API: {}",
response.error.message,
)),
_ => Err(anyhow!(
"Failed to connect to OpenAI API: {} {}",
response.status(),
body,
)),
}
}
}
#[derive(Copy, Clone, Serialize, Deserialize)]
pub enum OpenAiEmbeddingModel {
#[serde(rename = "text-embedding-3-small")]
TextEmbedding3Small,
#[serde(rename = "text-embedding-3-large")]
TextEmbedding3Large,
}
#[derive(Serialize)]
struct OpenAiEmbeddingRequest<'a> {
model: OpenAiEmbeddingModel,
input: Vec<&'a str>,
}
#[derive(Deserialize)]
pub struct OpenAiEmbeddingResponse {
pub data: Vec<OpenAiEmbedding>,
}
#[derive(Deserialize)]
pub struct OpenAiEmbedding {
pub embedding: Vec<f32>,
}
pub fn embed<'a>(
client: &dyn HttpClient,
api_url: &str,
api_key: &str,
model: OpenAiEmbeddingModel,
texts: impl IntoIterator<Item = &'a str>,
) -> impl 'static + Future<Output = Result<OpenAiEmbeddingResponse>> {
let uri = format!("{api_url}/embeddings");
let request = OpenAiEmbeddingRequest {
model,
input: texts.into_iter().collect(),
};
let body = AsyncBody::from(serde_json::to_string(&request).unwrap());
let request = HttpRequest::builder()
.method(Method::POST)
.uri(uri)
.header("Content-Type", "application/json")
.header("Authorization", format!("Bearer {}", api_key))
.body(body)
.map(|request| client.send(request));
async move {
let mut response = request?.await?;
let mut body = String::new();
response.body_mut().read_to_string(&mut body).await?;
if response.status().is_success() {
let response: OpenAiEmbeddingResponse =
serde_json::from_str(&body).context("failed to parse OpenAI embedding response")?;
Ok(response)
} else {
Err(anyhow!(
"error during embedding, status: {:?}, body: {:?}",
response.status(),
body
))
}
}
}
pub async fn extract_tool_args_from_events(
tool_name: String,
mut events: Pin<Box<dyn Send + Stream<Item = Result<ResponseStreamEvent>>>>,
) -> Result<impl Send + Stream<Item = Result<String>>> {
let mut tool_use_index = None;
let mut first_chunk = None;
while let Some(event) = events.next().await {
let call = event?.choices.into_iter().find_map(|choice| {
choice.delta.tool_calls?.into_iter().find_map(|call| {
if call.function.as_ref()?.name.as_deref()? == tool_name {
Some(call)
} else {
None
}
})
});
if let Some(call) = call {
tool_use_index = Some(call.index);
first_chunk = call.function.and_then(|func| func.arguments);
break;
}
}
let Some(tool_use_index) = tool_use_index else {
return Err(anyhow!("tool not used"));
};
Ok(events.filter_map(move |event| {
let result = match event {
Err(error) => Some(Err(error)),
Ok(ResponseStreamEvent { choices, .. }) => choices.into_iter().find_map(|choice| {
choice.delta.tool_calls?.into_iter().find_map(|call| {
if call.index == tool_use_index {
let func = call.function?;
let mut arguments = func.arguments?;
if let Some(mut first_chunk) = first_chunk.take() {
first_chunk.push_str(&arguments);
arguments = first_chunk
}
Some(Ok(arguments))
} else {
None
}
})
}),
};
async move { result }
}))
}
pub fn extract_text_from_events(
response: impl Stream<Item = Result<ResponseStreamEvent>>,
) -> impl Stream<Item = Result<String>> {
response.filter_map(|response| async move {
match response {
Ok(mut response) => Some(Ok(response.choices.pop()?.delta.content?)),
Err(error) => Some(Err(error)),
}
})
}