1 stable release

1.0.0 Aug 20, 2023

#12 in #conversation


Used in 2 crates

MIT license

11KB
224 lines

ChatGPT-rs

This library is an asynchronous Rust wrapper over the OpenAI ChatGPT API. It supports conversations, message persistence and ChatGPT functions.

Regarding ChatGPT Functions

The function API (available in v1.2.0+) is currently experimental and may not work as intended. If you encounter any issues or undefined behaviour, please, create an issue in this repository!

MSRV

The Minimum Supported Rust Version for this library is 1.71.1

Usage

Here is a simple usage of the API, getting completion for a single message. You can see more practical examples in the examples directory.

use chatgpt::prelude::*;
use chatgpt::types::{CompletionResponse};

#[tokio::main]
async fn main() -> Result<()> {
    // Getting the API key here
    let key = args().nth(1).unwrap();

    /// Creating a new ChatGPT client.
    /// Note that it requires an API key, and uses
    /// tokens from your OpenAI API account balance.
    let client = ChatGPT::new(key)?;

    /// Sending a message and getting the completion
    let response: CompletionResponse = client
        .send_message("Describe in five words the Rust programming language.")
        .await?;

    println!("Response: {}", response.message().content);

    Ok(())
}

Streaming Responses

If you wish to gradually build the response message, you may use the streams feature (not enabled by default) of the crate, and special methods to request streamed responses.

Here is an example:

// Acquiring a streamed response
// Note, that the `futures_util` crate is required for most
// stream related utility methods
let stream = client
    .send_message_streaming("Could you name me a few popular Rust backend server frameworks?")
    .await?;

// Iterating over stream contents
stream
    .for_each(|each| async move {
        match each {
            ResponseChunk::Content {
                delta,
                response_index: _,
            } => {
                // Printing part of response without the newline
                print!("{delta}");
                // Manually flushing the standard output, as `print` macro does not do that
                stdout().lock().flush().unwrap();
            }
            _ => {}
        }
    })
    .await;
}

Note that the returned streams normally don't have any utility methods, so you will have to use a StreamExt method from your async library of choice (e.g. futures-util or tokio).

Conversations

Conversations are the threads in which ChatGPT can analyze previous messages and chain it's thoughts. They also automatically store all the message history.

Here is an example:

// Creating a new conversation
let mut conversation: Conversation = client.new_conversation();

// Sending messages to the conversation
let response_a: CompletionResponse = conversation
    .send_message("Could you describe the Rust programming language in 5 words?")
    .await?;
let response_b: CompletionResponse = conversation
    .send_message("Now could you do the same, but for Kotlin?")
    .await?;

// You can also access the message history itself
for message in &conversation.history {
    println!("{message:#?}")
}

This way of creating a conversation creates it with the default introductory message, which roughly is: You are ChatGPT, an AI model developed by OpenAI. Answer as concisely as possible. Today is: {today's date}.

However, you can specify the introductory message yourself this way:

let mut conversation: Conversation = client.new_conversation_directed("You are RustGPT, when answering any questions, you always shift the topic of the conversation to the Rust programming language.");
// Continue with the new conversation

Conversation Streaming

Conversations also support returning streamed responses (with the streams feature).

NOTE: Streamed responses do not automatically save returned message to history, so you will have to do it manually by yourself.

Here is an example:

// Acquiring a streamed response
// Note, that the `futures_util` crate is required for most
// stream related utility methods
let mut stream = conversation
    .send_message_streaming("Could you name me a few popular Rust backend server frameworks?")
    .await?;

    // Iterating over a stream and collecting the results into a vector
let mut output: Vec<ResponseChunk> = Vec::new();
while let Some(chunk) = stream.next().await {
    match chunk {
        ResponseChunk::Content {
            delta,
            response_index,
        } => {
            // Printing part of response without the newline
            print!("{delta}");
            // Manually flushing the standard output, as `print` macro does not do that
            stdout().lock().flush().unwrap();
            output.push(ResponseChunk::Content {
                delta,
                response_index,
            });
        }
        // We don't really care about other types, other than parsing them into a ChatMessage later
        other => output.push(other),
    }
}

// Parsing ChatMessage from the response chunks and saving it to the conversation history
let messages = ChatMessage::from_response_chunks(output);
conversation.history.push(messages[0].to_owned());

Function Calls

ChatGPT-rs supports function calling API. Requires the functions feature.

You can define functions with the gpt_function attribute macro, like this:

use chatgpt::prelude::*;

/// Says hello to a user
/// 
/// * user_name - Name of the user to greet
#[gpt_function]
async fn say_hello(user_name: String) {
    println!("Hello, {user_name}!")
}

// ... within your conversation, before sending first message
let mut conversation = client.new_conversation();
// note that you need to call the function when adding it
conversation.add_function(say_hello());
let response = conversation
    .send_message_functions("Could you greet user with name `maxus`?")
    .await?;
// At this point, if function call was issued it was already processed
// and subsequent response was sent

As you can see, GPT functions must have a description so the model knows when to call them and what they do. In ChatGPT-rs function descriptions are represented as simple rust docs. Each argument is documented as * {argument name} - {argument description}. Function arguments are processed from JSON, so as long as they implement schemars::JsonSchema and serde::Deserialize they will be parsed correctly.

By default, ChatGPT-rs uses minimal schemars features, enable feature functions_extra to add support for uuid, chrono, url and either, or define your own structure and derive schemars::JsonSchema and serde::Deserialize:

use schemars::JsonSchema;
use serde::Deserialize;

#[derive(JsonSchema, Deserialize)]
struct Args {
    /// Name of the user
    user_name: String,
    /// New age of the user
    user_age: u16
}

/// Wishes happy birthday to the user
/// 
/// * args - Arguments
#[gpt_function]
async fn happy_birthday(args: Args) {
    println!("Hello, {}, You are now {}!", args.user_name, args.user_age);
}

Functions can also return any data (as long as it implements serde::Serialize) and it will be returned to the model.

/// Does some heavy computations and returns result
/// 
/// * input - Input data as vector of floats
#[gpt_function]
async fn do_heavy_computation(input: Vec<f64>) -> Vec<f64> {
    let output: Vec<f64> = // ... Do something with the input ...  
    return output;
}

By default, functions are only sent to API by calling the send_message_functions method. If you wish to enable automatic function sending with each message, you can set the always_send_functions property within Conversation to true.

Current function limitations are:

  • They must be async.
  • Since they are counted as tokens, you might want to limit function sending and/or their description length.

Function Call Validation

As stated in the official ChatGPT documentation, ChatGPT may hallucinate nonexistent functions or provide invalid JSON. To mitigate it, ChatGPT-rs provides FunctionValidationStrategy. If set to Strict within the client model configuration, a system message will be sent to the model correcting it whenever it fails to call function correctly.

Conversation Persistence

You can currently store the conversation's message in two formats: JSON or postcard. They can be toggled on or off using the json and postcard features respectively.

Since the ChatMessage struct derives serde's Serialize and Deserialize traits, you can also use any serde-compatible serialization library, as the history field and the Conversation::new_with_history() method are public in the Conversation struct.

Persistence with JSON

Requires the json feature (enabled by default)

// Create a new conversation here
let mut conversation: Conversation = ...;

// ... send messages to the conversation ...

// Saving the conversation
conversation.save_history_json("my-conversation.json").await?;

// You can later read this conversation history again
let mut restored = client
    .restore_conversation_json("my-conversation.json")
    .await?;

Persistence with Postcard

Requires the postcard feature (disabled by default)

// Create a new conversation here
let mut conversation: Conversation = ...;

// ... send messages to the conversation ...

// Saving the conversation
conversation.save_history_postcard("my-conversation.bin").await?;

// You can later read this conversation history again
let mut restored = client
    .restore_conversation_postcard("my-conversation.bin")
    .await?;

Advanced configuration

You can configure your model further with ModelConfigurationBuilder, which also allows to use proxies:

// Getting the API key here
let key = args().nth(1).unwrap();

// Creating a new ChatGPT client with extra settings.
// Note that it might not require an API key depending on proxy
let client = ChatGPT::new_with_config(
    key,
    ModelConfigurationBuilder::default()
        .api_url("https://api.pawan.krd/v1/chat/completions")
        .temperature(1.0)
        .engine(ChatGPTEngine::Gpt4_32k)
        .build()
        .unwrap(),
)?;

Dependencies

~0.6–1MB
~23K SLoC