rig/providers/

perplexity.rs

//! Perplexity API client and Rig integration
//!
//! # Example
//! ```
//! use rig::providers::perplexity;
//!
//! let client = perplexity::Client::new("YOUR_API_KEY");
//!
//! let llama_3_1_sonar_small_online = client.completion_model(perplexity::LLAMA_3_1_SONAR_SMALL_ONLINE);
//! ```
use crate::client::BearerAuth;
use crate::completion::CompletionRequest;
use crate::providers::openai;
use crate::providers::openai::send_compatible_streaming_request;
use crate::streaming::StreamingCompletionResponse;
use crate::{
    OneOrMany,
    client::{
        self, Capabilities, Capable, DebugExt, Nothing, Provider, ProviderBuilder, ProviderClient,
    },
    completion::{self, CompletionError, MessageError, message},
    http_client::{self, HttpClientExt},
};
use bytes::Bytes;
use serde::{Deserialize, Serialize};
use tracing::{Instrument, info_span};

// ================================================================
// Main Cohere Client
// ================================================================
const PERPLEXITY_API_BASE_URL: &str = "https://api.perplexity.ai";

#[derive(Debug, Default, Clone, Copy)]
pub struct PerplexityExt;

#[derive(Debug, Default, Clone, Copy)]
pub struct PerplexityBuilder;

type PerplexityApiKey = BearerAuth;

impl Provider for PerplexityExt {
    type Builder = PerplexityBuilder;

    // There is currently no way to verify a perplexity api key without consuming tokens
    const VERIFY_PATH: &'static str = "";

    fn build<H>(
        _: &crate::client::ClientBuilder<
            Self::Builder,
            <Self::Builder as crate::client::ProviderBuilder>::ApiKey,
            H,
        >,
    ) -> http_client::Result<Self> {
        Ok(Self)
    }
}

impl<H> Capabilities<H> for PerplexityExt {
    type Completion = Capable<CompletionModel<H>>;
    type Transcription = Nothing;
    type Embeddings = Nothing;
    #[cfg(feature = "image")]
    type ImageGeneration = Nothing;

    #[cfg(feature = "audio")]
    type AudioGeneration = Nothing;
}

impl DebugExt for PerplexityExt {}

impl ProviderBuilder for PerplexityBuilder {
    type Output = PerplexityExt;
    type ApiKey = PerplexityApiKey;

    const BASE_URL: &'static str = PERPLEXITY_API_BASE_URL;
}

pub type Client<H = reqwest::Client> = client::Client<PerplexityExt, H>;
pub type ClientBuilder<H = reqwest::Client> =
    client::ClientBuilder<PerplexityBuilder, PerplexityApiKey, H>;

impl ProviderClient for Client {
    type Input = String;

    /// Create a new Perplexity client from the `PERPLEXITY_API_KEY` environment variable.
    /// Panics if the environment variable is not set.
    fn from_env() -> Self {
        let api_key = std::env::var("PERPLEXITY_API_KEY").expect("PERPLEXITY_API_KEY not set");
        Self::new(&api_key).unwrap()
    }

    fn from_val(input: Self::Input) -> Self {
        Self::new(&input).unwrap()
    }
}

#[derive(Debug, Deserialize)]
struct ApiErrorResponse {
    message: String,
}

#[derive(Debug, Deserialize)]
#[serde(untagged)]
enum ApiResponse<T> {
    Ok(T),
    Err(ApiErrorResponse),
}

// ================================================================
// Perplexity Completion API
// ================================================================

pub const SONAR_PRO: &str = "sonar_pro";
pub const SONAR: &str = "sonar";

#[derive(Debug, Deserialize, Serialize)]
pub struct CompletionResponse {
    pub id: String,
    pub model: String,
    pub object: String,
    pub created: u64,
    #[serde(default)]
    pub choices: Vec<Choice>,
    pub usage: Usage,
}

#[derive(Serialize, Deserialize, Debug, Clone, PartialEq)]
pub struct Message {
    pub role: Role,
    pub content: String,
}

#[derive(Serialize, Deserialize, Debug, Clone, PartialEq)]
#[serde(rename_all = "lowercase")]
pub enum Role {
    System,
    User,
    Assistant,
}

#[derive(Deserialize, Debug, Serialize)]
pub struct Delta {
    pub role: Role,
    pub content: String,
}

#[derive(Deserialize, Debug, Serialize)]
pub struct Choice {
    pub index: usize,
    pub finish_reason: String,
    pub message: Message,
    pub delta: Delta,
}

#[derive(Deserialize, Debug, Serialize)]
pub struct Usage {
    pub prompt_tokens: u32,
    pub completion_tokens: u32,
    pub total_tokens: u32,
}

impl std::fmt::Display for Usage {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        write!(
            f,
            "Prompt tokens: {}\nCompletion tokens: {} Total tokens: {}",
            self.prompt_tokens, self.completion_tokens, self.total_tokens
        )
    }
}

impl TryFrom<CompletionResponse> for completion::CompletionResponse<CompletionResponse> {
    type Error = CompletionError;

    fn try_from(response: CompletionResponse) -> Result<Self, Self::Error> {
        let choice = response.choices.first().ok_or_else(|| {
            CompletionError::ResponseError("Response contained no choices".to_owned())
        })?;

        match &choice.message {
            Message {
                role: Role::Assistant,
                content,
            } => Ok(completion::CompletionResponse {
                choice: OneOrMany::one(content.clone().into()),
                usage: completion::Usage {
                    input_tokens: response.usage.prompt_tokens as u64,
                    output_tokens: response.usage.completion_tokens as u64,
                    total_tokens: response.usage.total_tokens as u64,
                },
                raw_response: response,
            }),
            _ => Err(CompletionError::ResponseError(
                "Response contained no assistant message".to_owned(),
            )),
        }
    }
}

#[derive(Debug, Serialize, Deserialize)]
pub(super) struct PerplexityCompletionRequest {
    model: String,
    pub messages: Vec<Message>,
    #[serde(skip_serializing_if = "Option::is_none")]
    pub temperature: Option<f64>,
    #[serde(skip_serializing_if = "Option::is_none")]
    pub max_tokens: Option<u64>,
    #[serde(flatten, skip_serializing_if = "Option::is_none")]
    additional_params: Option<serde_json::Value>,
    pub stream: bool,
}

impl TryFrom<(&str, CompletionRequest)> for PerplexityCompletionRequest {
    type Error = CompletionError;

    fn try_from((model, req): (&str, CompletionRequest)) -> Result<Self, Self::Error> {
        let mut partial_history = vec![];
        if let Some(docs) = req.normalized_documents() {
            partial_history.push(docs);
        }
        partial_history.extend(req.chat_history);

        // Initialize full history with preamble (or empty if non-existent)
        let mut full_history: Vec<Message> = req.preamble.map_or_else(Vec::new, |preamble| {
            vec![Message {
                role: Role::System,
                content: preamble,
            }]
        });

        // Convert and extend the rest of the history
        full_history.extend(
            partial_history
                .into_iter()
                .map(message::Message::try_into)
                .collect::<Result<Vec<Message>, _>>()?,
        );

        Ok(Self {
            model: model.to_string(),
            messages: full_history,
            temperature: req.temperature,
            max_tokens: req.max_tokens,
            additional_params: req.additional_params,
            stream: false,
        })
    }
}

#[derive(Clone)]
pub struct CompletionModel<T = reqwest::Client> {
    client: Client<T>,
    pub model: String,
}

impl<T> CompletionModel<T> {
    pub fn new(client: Client<T>, model: impl Into<String>) -> Self {
        Self {
            client,
            model: model.into(),
        }
    }
}

impl TryFrom<message::Message> for Message {
    type Error = MessageError;

    fn try_from(message: message::Message) -> Result<Self, Self::Error> {
        Ok(match message {
            message::Message::User { content } => {
                let collapsed_content = content
                    .into_iter()
                    .map(|content| match content {
                        message::UserContent::Text(message::Text { text }) => Ok(text),
                        _ => Err(MessageError::ConversionError(
                            "Only text content is supported by Perplexity".to_owned(),
                        )),
                    })
                    .collect::<Result<Vec<_>, _>>()?
                    .join("\n");

                Message {
                    role: Role::User,
                    content: collapsed_content,
                }
            }

            message::Message::Assistant { content, .. } => {
                let collapsed_content = content
                    .into_iter()
                    .map(|content| {
                        Ok(match content {
                            message::AssistantContent::Text(message::Text { text }) => text,
                            _ => return Err(MessageError::ConversionError(
                                "Only text assistant message content is supported by Perplexity"
                                    .to_owned(),
                            )),
                        })
                    })
                    .collect::<Result<Vec<_>, _>>()?
                    .join("\n");

                Message {
                    role: Role::Assistant,
                    content: collapsed_content,
                }
            }
        })
    }
}

impl From<Message> for message::Message {
    fn from(message: Message) -> Self {
        match message.role {
            Role::User => message::Message::user(message.content),
            Role::Assistant => message::Message::assistant(message.content),

            // System messages get coerced into user messages for ease of error handling.
            // They should be handled on the outside of `Message` conversions via the preamble.
            Role::System => message::Message::user(message.content),
        }
    }
}

impl<T> completion::CompletionModel for CompletionModel<T>
where
    T: HttpClientExt + Clone + Default + std::fmt::Debug + Send + 'static,
{
    type Response = CompletionResponse;
    type StreamingResponse = openai::StreamingCompletionResponse;

    type Client = Client<T>;

    fn make(client: &Self::Client, model: impl Into<String>) -> Self {
        Self::new(client.clone(), model)
    }

    async fn completion(
        &self,
        completion_request: completion::CompletionRequest,
    ) -> Result<completion::CompletionResponse<CompletionResponse>, CompletionError> {
        let span = if tracing::Span::current().is_disabled() {
            info_span!(
                target: "rig::completions",
                "chat",
                gen_ai.operation.name = "chat",
                gen_ai.provider.name = "perplexity",
                gen_ai.request.model = self.model,
                gen_ai.system_instructions = tracing::field::Empty,
                gen_ai.response.id = tracing::field::Empty,
                gen_ai.response.model = tracing::field::Empty,
                gen_ai.usage.output_tokens = tracing::field::Empty,
                gen_ai.usage.input_tokens = tracing::field::Empty,
            )
        } else {
            tracing::Span::current()
        };

        span.record("gen_ai.system_instructions", &completion_request.preamble);

        if completion_request.tool_choice.is_some() {
            tracing::warn!("WARNING: `tool_choice` not supported on Perplexity");
        }

        if !completion_request.tools.is_empty() {
            tracing::warn!("WARNING: `tools` not supported on Perplexity");
        }
        let request =
            PerplexityCompletionRequest::try_from((self.model.as_ref(), completion_request))?;

        if tracing::enabled!(tracing::Level::TRACE) {
            tracing::trace!(target: "rig::completions",
                "Perplexity completion request: {}",
                serde_json::to_string_pretty(&request)?
            );
        }

        let body = serde_json::to_vec(&request)?;

        let req = self
            .client
            .post("/v1/chat/completions")?
            .body(body)
            .map_err(http_client::Error::from)?;

        let async_block = async move {
            let response = self.client.send::<_, Bytes>(req).await?;

            let status = response.status();
            let response_body = response.into_body().into_future().await?.to_vec();

            if status.is_success() {
                match serde_json::from_slice::<ApiResponse<CompletionResponse>>(&response_body)? {
                    ApiResponse::Ok(response) => {
                        let span = tracing::Span::current();
                        span.record("gen_ai.usage.input_tokens", response.usage.prompt_tokens);
                        span.record(
                            "gen_ai.usage.output_tokens",
                            response.usage.completion_tokens,
                        );
                        span.record("gen_ai.response.id", response.id.to_string());
                        span.record("gen_ai.response.model_name", response.model.to_string());
                        if tracing::enabled!(tracing::Level::TRACE) {
                            tracing::trace!(target: "rig::responses",
                                "Perplexity completion response: {}",
                                serde_json::to_string_pretty(&response)?
                            );
                        }
                        Ok(response.try_into()?)
                    }
                    ApiResponse::Err(error) => Err(CompletionError::ProviderError(error.message)),
                }
            } else {
                Err(CompletionError::ProviderError(
                    String::from_utf8_lossy(&response_body).to_string(),
                ))
            }
        };

        async_block.instrument(span).await
    }

    async fn stream(
        &self,
        completion_request: completion::CompletionRequest,
    ) -> Result<StreamingCompletionResponse<Self::StreamingResponse>, CompletionError> {
        let span = if tracing::Span::current().is_disabled() {
            info_span!(
                target: "rig::completions",
                "chat_streaming",
                gen_ai.operation.name = "chat_streaming",
                gen_ai.provider.name = "perplexity",
                gen_ai.request.model = self.model,
                gen_ai.system_instructions = tracing::field::Empty,
                gen_ai.response.id = tracing::field::Empty,
                gen_ai.response.model = tracing::field::Empty,
                gen_ai.usage.output_tokens = tracing::field::Empty,
                gen_ai.usage.input_tokens = tracing::field::Empty,
            )
        } else {
            tracing::Span::current()
        };

        span.record("gen_ai.system_instructions", &completion_request.preamble);

        if completion_request.tool_choice.is_some() {
            tracing::warn!("WARNING: `tool_choice` not supported on Perplexity");
        }

        if !completion_request.tools.is_empty() {
            tracing::warn!("WARNING: `tools` not supported on Perplexity");
        }

        let mut request =
            PerplexityCompletionRequest::try_from((self.model.as_ref(), completion_request))?;
        request.stream = true;

        if tracing::enabled!(tracing::Level::TRACE) {
            tracing::trace!(target: "rig::completions",
                "Perplexity streaming completion request: {}",
                serde_json::to_string_pretty(&request)?
            );
        }

        let body = serde_json::to_vec(&request)?;

        let req = self
            .client
            .post("/chat/completions")?
            .body(body)
            .map_err(http_client::Error::from)?;

        send_compatible_streaming_request(self.client.clone(), req)
            .instrument(span)
            .await
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_deserialize_message() {
        let json_data = r#"
        {
            "role": "user",
            "content": "Hello, how can I help you?"
        }
        "#;

        let message: Message = serde_json::from_str(json_data).unwrap();
        assert_eq!(message.role, Role::User);
        assert_eq!(message.content, "Hello, how can I help you?");
    }

    #[test]
    fn test_serialize_message() {
        let message = Message {
            role: Role::Assistant,
            content: "I am here to assist you.".to_string(),
        };

        let json_data = serde_json::to_string(&message).unwrap();
        let expected_json = r#"{"role":"assistant","content":"I am here to assist you."}"#;
        assert_eq!(json_data, expected_json);
    }

    #[test]
    fn test_message_to_message_conversion() {
        let user_message = message::Message::user("User message");
        let assistant_message = message::Message::assistant("Assistant message");

        let converted_user_message: Message = user_message.clone().try_into().unwrap();
        let converted_assistant_message: Message = assistant_message.clone().try_into().unwrap();

        assert_eq!(converted_user_message.role, Role::User);
        assert_eq!(converted_user_message.content, "User message");

        assert_eq!(converted_assistant_message.role, Role::Assistant);
        assert_eq!(converted_assistant_message.content, "Assistant message");

        let back_to_user_message: message::Message = converted_user_message.into();
        let back_to_assistant_message: message::Message = converted_assistant_message.into();

        assert_eq!(user_message, back_to_user_message);
        assert_eq!(assistant_message, back_to_assistant_message);
    }
}