rig/providers/

deepseek.rs

//! DeepSeek API client and Rig integration
//!
//! # Example
//! ```
//! use rig::providers::deepseek;
//!
//! let client = deepseek::Client::new("DEEPSEEK_API_KEY");
//!
//! let deepseek_chat = client.completion_model(deepseek::DEEPSEEK_CHAT);
//! ```

use crate::json_utils::empty_or_none;
use async_stream::stream;
use bytes::Bytes;
use futures::StreamExt;
use http::Request;
use std::collections::HashMap;
use tracing::{Instrument, Level, enabled, info_span};

use crate::client::{
    self, BearerAuth, Capabilities, Capable, DebugExt, Nothing, Provider, ProviderBuilder,
    ProviderClient,
};
use crate::completion::GetTokenUsage;
use crate::http_client::sse::{Event, GenericEventSource};
use crate::http_client::{self, HttpClientExt};
use crate::message::{Document, DocumentSourceKind};
use crate::{
    OneOrMany,
    completion::{self, CompletionError, CompletionRequest},
    json_utils, message,
};
use serde::{Deserialize, Serialize};

use super::openai::StreamingToolCall;

// ================================================================
// Main DeepSeek Client
// ================================================================
const DEEPSEEK_API_BASE_URL: &str = "https://api.deepseek.com";

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

type DeepSeekApiKey = BearerAuth;

impl Provider for DeepSeekExt {
    type Builder = DeepSeekExtBuilder;

    const VERIFY_PATH: &'static str = "/user/balance";

    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 DeepSeekExt {
    type Completion = Capable<CompletionModel<H>>;
    type Embeddings = Nothing;
    type Transcription = Nothing;
    #[cfg(feature = "image")]
    type ImageGeneration = Nothing;
    #[cfg(feature = "audio")]
    type AudioGeneration = Nothing;
}

impl DebugExt for DeepSeekExt {}

impl ProviderBuilder for DeepSeekExtBuilder {
    type Output = DeepSeekExt;
    type ApiKey = DeepSeekApiKey;

    const BASE_URL: &'static str = DEEPSEEK_API_BASE_URL;
}

pub type Client<H = reqwest::Client> = client::Client<DeepSeekExt, H>;
pub type ClientBuilder<H = reqwest::Client> = client::ClientBuilder<DeepSeekExtBuilder, String, H>;

impl ProviderClient for Client {
    type Input = DeepSeekApiKey;

    // If you prefer the environment variable approach:
    fn from_env() -> Self {
        let api_key = std::env::var("DEEPSEEK_API_KEY").expect("DEEPSEEK_API_KEY not set");
        let mut client_builder = Self::builder();
        client_builder.headers_mut().insert(
            http::header::CONTENT_TYPE,
            http::HeaderValue::from_static("application/json"),
        );
        let client_builder = client_builder.api_key(&api_key);
        client_builder.build().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),
}

impl From<ApiErrorResponse> for CompletionError {
    fn from(err: ApiErrorResponse) -> Self {
        CompletionError::ProviderError(err.message)
    }
}

/// The response shape from the DeepSeek API
#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct CompletionResponse {
    // We'll match the JSON:
    pub choices: Vec<Choice>,
    pub usage: Usage,
    // you may want other fields
}

#[derive(Clone, Debug, Serialize, Deserialize, Default)]
pub struct Usage {
    pub completion_tokens: u32,
    pub prompt_tokens: u32,
    pub prompt_cache_hit_tokens: u32,
    pub prompt_cache_miss_tokens: u32,
    pub total_tokens: u32,
    #[serde(skip_serializing_if = "Option::is_none")]
    pub completion_tokens_details: Option<CompletionTokensDetails>,
    #[serde(skip_serializing_if = "Option::is_none")]
    pub prompt_tokens_details: Option<PromptTokensDetails>,
}

impl Usage {
    fn new() -> Self {
        Self {
            completion_tokens: 0,
            prompt_tokens: 0,
            prompt_cache_hit_tokens: 0,
            prompt_cache_miss_tokens: 0,
            total_tokens: 0,
            completion_tokens_details: None,
            prompt_tokens_details: None,
        }
    }
}

#[derive(Clone, Debug, Serialize, Deserialize, Default)]
pub struct CompletionTokensDetails {
    #[serde(skip_serializing_if = "Option::is_none")]
    pub reasoning_tokens: Option<u32>,
}

#[derive(Clone, Debug, Serialize, Deserialize, Default)]
pub struct PromptTokensDetails {
    #[serde(skip_serializing_if = "Option::is_none")]
    pub cached_tokens: Option<u32>,
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
pub struct Choice {
    pub index: usize,
    pub message: Message,
    pub logprobs: Option<serde_json::Value>,
    pub finish_reason: String,
}

#[derive(Debug, Serialize, Deserialize, PartialEq, Clone)]
#[serde(tag = "role", rename_all = "lowercase")]
pub enum Message {
    System {
        content: String,
        #[serde(skip_serializing_if = "Option::is_none")]
        name: Option<String>,
    },
    User {
        content: String,
        #[serde(skip_serializing_if = "Option::is_none")]
        name: Option<String>,
    },
    Assistant {
        content: String,
        #[serde(skip_serializing_if = "Option::is_none")]
        name: Option<String>,
        #[serde(
            default,
            deserialize_with = "json_utils::null_or_vec",
            skip_serializing_if = "Vec::is_empty"
        )]
        tool_calls: Vec<ToolCall>,
        /// only exists on `deepseek-reasoner` model at time of addition
        #[serde(skip_serializing_if = "Option::is_none")]
        reasoning_content: Option<String>,
    },
    #[serde(rename = "tool")]
    ToolResult {
        tool_call_id: String,
        content: String,
    },
}

impl Message {
    pub fn system(content: &str) -> Self {
        Message::System {
            content: content.to_owned(),
            name: None,
        }
    }
}

impl From<message::ToolResult> for Message {
    fn from(tool_result: message::ToolResult) -> Self {
        let content = match tool_result.content.first() {
            message::ToolResultContent::Text(text) => text.text,
            message::ToolResultContent::Image(_) => String::from("[Image]"),
        };

        Message::ToolResult {
            tool_call_id: tool_result.id,
            content,
        }
    }
}

impl From<message::ToolCall> for ToolCall {
    fn from(tool_call: message::ToolCall) -> Self {
        Self {
            id: tool_call.id,
            // TODO: update index when we have it
            index: 0,
            r#type: ToolType::Function,
            function: Function {
                name: tool_call.function.name,
                arguments: tool_call.function.arguments,
            },
        }
    }
}

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

    fn try_from(message: message::Message) -> Result<Self, Self::Error> {
        match message {
            message::Message::User { content } => {
                // extract tool results
                let mut messages = vec![];

                let tool_results = content
                    .clone()
                    .into_iter()
                    .filter_map(|content| match content {
                        message::UserContent::ToolResult(tool_result) => {
                            Some(Message::from(tool_result))
                        }
                        _ => None,
                    })
                    .collect::<Vec<_>>();

                messages.extend(tool_results);

                // extract text results
                let text_messages = content
                    .into_iter()
                    .filter_map(|content| match content {
                        message::UserContent::Text(text) => Some(Message::User {
                            content: text.text,
                            name: None,
                        }),
                        message::UserContent::Document(Document {
                            data:
                                DocumentSourceKind::Base64(content)
                                | DocumentSourceKind::String(content),
                            ..
                        }) => Some(Message::User {
                            content,
                            name: None,
                        }),
                        _ => None,
                    })
                    .collect::<Vec<_>>();
                messages.extend(text_messages);

                Ok(messages)
            }
            message::Message::Assistant { content, .. } => {
                let mut messages: Vec<Message> = vec![];
                let mut text_content = String::new();
                let mut reasoning_content = String::new();

                content.iter().for_each(|content| match content {
                    message::AssistantContent::Text(text) => {
                        text_content.push_str(text.text());
                    }
                    message::AssistantContent::Reasoning(reasoning) => {
                        reasoning_content.push_str(&reasoning.reasoning.join("\n"));
                    }
                    _ => {}
                });

                messages.push(Message::Assistant {
                    content: text_content,
                    name: None,
                    tool_calls: vec![],
                    reasoning_content: if reasoning_content.is_empty() {
                        None
                    } else {
                        Some(reasoning_content)
                    },
                });

                // extract tool calls
                let tool_calls = content
                    .clone()
                    .into_iter()
                    .filter_map(|content| match content {
                        message::AssistantContent::ToolCall(tool_call) => {
                            Some(ToolCall::from(tool_call))
                        }
                        _ => None,
                    })
                    .collect::<Vec<_>>();

                // if we have tool calls, we add a new Assistant message with them
                if !tool_calls.is_empty() {
                    messages.push(Message::Assistant {
                        content: "".to_string(),
                        name: None,
                        tool_calls,
                        reasoning_content: None,
                    });
                }

                Ok(messages)
            }
        }
    }
}

#[derive(Debug, Serialize, Deserialize, PartialEq, Clone)]
pub struct ToolCall {
    pub id: String,
    pub index: usize,
    #[serde(default)]
    pub r#type: ToolType,
    pub function: Function,
}

#[derive(Debug, Serialize, Deserialize, PartialEq, Clone)]
pub struct Function {
    pub name: String,
    #[serde(with = "json_utils::stringified_json")]
    pub arguments: serde_json::Value,
}

#[derive(Default, Debug, Serialize, Deserialize, PartialEq, Clone)]
#[serde(rename_all = "lowercase")]
pub enum ToolType {
    #[default]
    Function,
}

#[derive(Clone, Debug, Deserialize, Serialize)]
pub struct ToolDefinition {
    pub r#type: String,
    pub function: completion::ToolDefinition,
}

impl From<crate::completion::ToolDefinition> for ToolDefinition {
    fn from(tool: crate::completion::ToolDefinition) -> Self {
        Self {
            r#type: "function".into(),
            function: tool,
        }
    }
}

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())
        })?;
        let content = match &choice.message {
            Message::Assistant {
                content,
                tool_calls,
                reasoning_content,
                ..
            } => {
                let mut content = if content.trim().is_empty() {
                    vec![]
                } else {
                    vec![completion::AssistantContent::text(content)]
                };

                content.extend(
                    tool_calls
                        .iter()
                        .map(|call| {
                            completion::AssistantContent::tool_call(
                                &call.id,
                                &call.function.name,
                                call.function.arguments.clone(),
                            )
                        })
                        .collect::<Vec<_>>(),
                );

                if let Some(reasoning_content) = reasoning_content {
                    content.push(completion::AssistantContent::reasoning(reasoning_content));
                }

                Ok(content)
            }
            _ => Err(CompletionError::ResponseError(
                "Response did not contain a valid message or tool call".into(),
            )),
        }?;

        let choice = OneOrMany::many(content).map_err(|_| {
            CompletionError::ResponseError(
                "Response contained no message or tool call (empty)".to_owned(),
            )
        })?;

        let 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,
            cached_input_tokens: response
                .usage
                .prompt_tokens_details
                .as_ref()
                .and_then(|d| d.cached_tokens)
                .map(|c| c as u64)
                .unwrap_or(0),
        };

        Ok(completion::CompletionResponse {
            choice,
            usage,
            raw_response: response,
        })
    }
}

#[derive(Debug, Serialize, Deserialize)]
pub(super) struct DeepseekCompletionRequest {
    model: String,
    pub messages: Vec<Message>,
    #[serde(skip_serializing_if = "Option::is_none")]
    temperature: Option<f64>,
    #[serde(skip_serializing_if = "Vec::is_empty")]
    tools: Vec<ToolDefinition>,
    #[serde(skip_serializing_if = "Option::is_none")]
    tool_choice: Option<crate::providers::openrouter::ToolChoice>,
    #[serde(flatten, skip_serializing_if = "Option::is_none")]
    pub additional_params: Option<serde_json::Value>,
}

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

    fn try_from((model, req): (&str, CompletionRequest)) -> Result<Self, Self::Error> {
        let mut full_history: Vec<Message> = match &req.preamble {
            Some(preamble) => vec![Message::system(preamble)],
            None => vec![],
        };

        if let Some(docs) = req.normalized_documents() {
            let docs: Vec<Message> = docs.try_into()?;
            full_history.extend(docs);
        }

        let chat_history: Vec<Message> = req
            .chat_history
            .clone()
            .into_iter()
            .map(|message| message.try_into())
            .collect::<Result<Vec<Vec<Message>>, _>>()?
            .into_iter()
            .flatten()
            .collect();

        let mut last_reasoning_content = None;
        let mut last_assistant_idx = None;

        for message in chat_history {
            if let Message::Assistant {
                reasoning_content, ..
            } = &message
            {
                if let Some(content) = reasoning_content {
                    last_reasoning_content = Some(content.clone());
                } else {
                    last_assistant_idx = Some(full_history.len());
                    full_history.push(message);
                }
            } else {
                full_history.push(message);
            }
        }

        // Merge last reasoning content into the last assistant message.
        // Note that we only need to preserve the last reasoning content.
        if let (Some(idx), Some(reasoning)) = (last_assistant_idx, last_reasoning_content)
            && let Message::Assistant {
                ref mut reasoning_content,
                ..
            } = full_history[idx]
        {
            *reasoning_content = Some(reasoning);
        }

        let tool_choice = req
            .tool_choice
            .clone()
            .map(crate::providers::openrouter::ToolChoice::try_from)
            .transpose()?;

        Ok(Self {
            model: model.to_string(),
            messages: full_history,
            temperature: req.temperature,
            tools: req
                .tools
                .clone()
                .into_iter()
                .map(ToolDefinition::from)
                .collect::<Vec<_>>(),
            tool_choice,
            additional_params: req.additional_params,
        })
    }
}

/// The struct implementing the `CompletionModel` trait
#[derive(Clone)]
pub struct CompletionModel<T = reqwest::Client> {
    pub client: Client<T>,
    pub model: String,
}

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

    type Client = Client<T>;

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

    async fn completion(
        &self,
        completion_request: CompletionRequest,
    ) -> Result<
        completion::CompletionResponse<CompletionResponse>,
        crate::completion::CompletionError,
    > {
        let span = if tracing::Span::current().is_disabled() {
            info_span!(
                target: "rig::completions",
                "chat",
                gen_ai.operation.name = "chat",
                gen_ai.provider.name = "deepseek",
                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);

        let request =
            DeepseekCompletionRequest::try_from((self.model.as_ref(), completion_request))?;

        if enabled!(Level::TRACE) {
            tracing::trace!(target: "rig::completions",
                "DeepSeek 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(|e| CompletionError::HttpError(e.into()))?;

        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,
                        );
                        if enabled!(Level::TRACE) {
                            tracing::trace!(target: "rig::completions",
                                "DeepSeek completion response: {}",
                                serde_json::to_string_pretty(&response)?
                            );
                        }
                        response.try_into()
                    }
                    ApiResponse::Err(err) => Err(CompletionError::ProviderError(err.message)),
                }
            } else {
                Err(CompletionError::ProviderError(
                    String::from_utf8_lossy(&response_body).to_string(),
                ))
            }
        }
        .instrument(span)
        .await
    }

    async fn stream(
        &self,
        completion_request: CompletionRequest,
    ) -> Result<
        crate::streaming::StreamingCompletionResponse<Self::StreamingResponse>,
        CompletionError,
    > {
        let preamble = completion_request.preamble.clone();
        let mut request =
            DeepseekCompletionRequest::try_from((self.model.as_ref(), completion_request))?;

        let params = json_utils::merge(
            request.additional_params.unwrap_or(serde_json::json!({})),
            serde_json::json!({"stream": true, "stream_options": {"include_usage": true} }),
        );

        request.additional_params = Some(params);

        if enabled!(Level::TRACE) {
            tracing::trace!(target: "rig::completions",
                "DeepSeek 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(|e| CompletionError::HttpError(e.into()))?;

        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 = "deepseek",
                gen_ai.request.model = self.model,
                gen_ai.system_instructions = preamble,
                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()
        };

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

#[derive(Deserialize, Debug)]
pub struct StreamingDelta {
    #[serde(default)]
    content: Option<String>,
    #[serde(default, deserialize_with = "json_utils::null_or_vec")]
    tool_calls: Vec<StreamingToolCall>,
    reasoning_content: Option<String>,
}

#[derive(Deserialize, Debug)]
struct StreamingChoice {
    delta: StreamingDelta,
}

#[derive(Deserialize, Debug)]
struct StreamingCompletionChunk {
    choices: Vec<StreamingChoice>,
    usage: Option<Usage>,
}

#[derive(Clone, Deserialize, Serialize, Debug)]
pub struct StreamingCompletionResponse {
    pub usage: Usage,
}

impl GetTokenUsage for StreamingCompletionResponse {
    fn token_usage(&self) -> Option<crate::completion::Usage> {
        let mut usage = crate::completion::Usage::new();
        usage.input_tokens = self.usage.prompt_tokens as u64;
        usage.output_tokens = self.usage.completion_tokens as u64;
        usage.total_tokens = self.usage.total_tokens as u64;
        usage.cached_input_tokens = self
            .usage
            .prompt_tokens_details
            .as_ref()
            .and_then(|d| d.cached_tokens)
            .map(|c| c as u64)
            .unwrap_or(0);

        Some(usage)
    }
}

pub async fn send_compatible_streaming_request<T>(
    http_client: T,
    req: Request<Vec<u8>>,
) -> Result<
    crate::streaming::StreamingCompletionResponse<StreamingCompletionResponse>,
    CompletionError,
>
where
    T: HttpClientExt + Clone + 'static,
{
    let mut event_source = GenericEventSource::new(http_client, req);

    let stream = stream! {
        let mut final_usage = Usage::new();
        let mut text_response = String::new();
        let mut calls: HashMap<usize, (String, String, String)> = HashMap::new();

        while let Some(event_result) = event_source.next().await {
            match event_result {
                Ok(Event::Open) => {
                    tracing::trace!("SSE connection opened");
                    continue;
                }
                Ok(Event::Message(message)) => {
                    if message.data.trim().is_empty() || message.data == "[DONE]" {
                        continue;
                    }

                    let parsed = serde_json::from_str::<StreamingCompletionChunk>(&message.data);
                    let Ok(data) = parsed else {
                        let err = parsed.unwrap_err();
                        tracing::debug!("Couldn't parse SSE payload as StreamingCompletionChunk: {:?}", err);
                        continue;
                    };

                    if let Some(choice) = data.choices.first() {
                        let delta = &choice.delta;

                        if !delta.tool_calls.is_empty() {
                            for tool_call in &delta.tool_calls {
                                let function = &tool_call.function;

                                // Start of tool call
                                if function.name.as_ref().map(|s| !s.is_empty()).unwrap_or(false)
                                    && empty_or_none(&function.arguments)
                                {
                                    let id = tool_call.id.clone().unwrap_or_default();
                                    let name = function.name.clone().unwrap();
                                    calls.insert(tool_call.index, (id, name, String::new()));
                                }
                                // Continuation of tool call
                                else if function.name.as_ref().map(|s| s.is_empty()).unwrap_or(true)
                                    && let Some(arguments) = &function.arguments
                                    && !arguments.is_empty()
                                {
                                    if let Some((id, name, existing_args)) = calls.get(&tool_call.index) {
                                        let combined = format!("{}{}", existing_args, arguments);
                                        calls.insert(tool_call.index, (id.clone(), name.clone(), combined));
                                    } else {
                                        tracing::debug!("Partial tool call received but tool call was never started.");
                                    }
                                }
                                // Complete tool call
                                else {
                                    let id = tool_call.id.clone().unwrap_or_default();
                                    let name = function.name.clone().unwrap_or_default();
                                    let arguments_str = function.arguments.clone().unwrap_or_default();

                                    let Ok(arguments_json) = serde_json::from_str::<serde_json::Value>(&arguments_str) else {
                                        tracing::debug!("Couldn't parse tool call args '{}'", arguments_str);
                                        continue;
                                    };

                                    yield Ok(crate::streaming::RawStreamingChoice::ToolCall(
                                        crate::streaming::RawStreamingToolCall::new(id, name, arguments_json)
                                    ));
                                }
                            }
                        }

                        // DeepSeek-specific reasoning stream
                        if let Some(content) = &delta.reasoning_content {
                            yield Ok(crate::streaming::RawStreamingChoice::ReasoningDelta {
                                id: None,
                                reasoning: content.to_string()
                            });
                        }

                        if let Some(content) = &delta.content {
                            text_response += content;
                            yield Ok(crate::streaming::RawStreamingChoice::Message(content.clone()));
                        }
                    }

                    if let Some(usage) = data.usage {
                        final_usage = usage.clone();
                    }
                }
                Err(crate::http_client::Error::StreamEnded) => {
                    break;
                }
                Err(err) => {
                    tracing::error!(?err, "SSE error");
                    yield Err(CompletionError::ResponseError(err.to_string()));
                    break;
                }
            }
        }

        event_source.close();

        let mut tool_calls = Vec::new();
        // Flush accumulated tool calls
        for (index, (id, name, arguments)) in calls {
            let Ok(arguments_json) = serde_json::from_str::<serde_json::Value>(&arguments) else {
                continue;
            };

            tool_calls.push(ToolCall {
                id: id.clone(),
                index,
                r#type: ToolType::Function,
                function: Function {
                    name: name.clone(),
                    arguments: arguments_json.clone()
                }
            });
            yield Ok(crate::streaming::RawStreamingChoice::ToolCall(
                crate::streaming::RawStreamingToolCall::new(id, name, arguments_json)
            ));
        }

        yield Ok(crate::streaming::RawStreamingChoice::FinalResponse(
            StreamingCompletionResponse { usage: final_usage.clone() }
        ));
    };

    Ok(crate::streaming::StreamingCompletionResponse::stream(
        Box::pin(stream),
    ))
}

// ================================================================
// DeepSeek Completion API
// ================================================================
pub const DEEPSEEK_CHAT: &str = "deepseek-chat";
pub const DEEPSEEK_REASONER: &str = "deepseek-reasoner";

// Tests
#[cfg(test)]
mod tests {

    use super::*;

    #[test]
    fn test_deserialize_vec_choice() {
        let data = r#"[{
            "finish_reason": "stop",
            "index": 0,
            "logprobs": null,
            "message":{"role":"assistant","content":"Hello, world!"}
            }]"#;

        let choices: Vec<Choice> = serde_json::from_str(data).unwrap();
        assert_eq!(choices.len(), 1);
        match &choices.first().unwrap().message {
            Message::Assistant { content, .. } => assert_eq!(content, "Hello, world!"),
            _ => panic!("Expected assistant message"),
        }
    }

    #[test]
    fn test_deserialize_deepseek_response() {
        let data = r#"{
            "choices":[{
                "finish_reason": "stop",
                "index": 0,
                "logprobs": null,
                "message":{"role":"assistant","content":"Hello, world!"}
            }],
            "usage": {
                "completion_tokens": 0,
                "prompt_tokens": 0,
                "prompt_cache_hit_tokens": 0,
                "prompt_cache_miss_tokens": 0,
                "total_tokens": 0
            }
        }"#;

        let jd = &mut serde_json::Deserializer::from_str(data);
        let result: Result<CompletionResponse, _> = serde_path_to_error::deserialize(jd);
        match result {
            Ok(response) => match &response.choices.first().unwrap().message {
                Message::Assistant { content, .. } => assert_eq!(content, "Hello, world!"),
                _ => panic!("Expected assistant message"),
            },
            Err(err) => {
                panic!("Deserialization error at {}: {}", err.path(), err);
            }
        }
    }

    #[test]
    fn test_deserialize_example_response() {
        let data = r#"
        {
            "id": "e45f6c68-9d9e-43de-beb4-4f402b850feb",
            "object": "chat.completion",
            "created": 0,
            "model": "deepseek-chat",
            "choices": [
                {
                    "index": 0,
                    "message": {
                        "role": "assistant",
                        "content": "Why don’t skeletons fight each other?  \nBecause they don’t have the guts! 😄"
                    },
                    "logprobs": null,
                    "finish_reason": "stop"
                }
            ],
            "usage": {
                "prompt_tokens": 13,
                "completion_tokens": 32,
                "total_tokens": 45,
                "prompt_tokens_details": {
                    "cached_tokens": 0
                },
                "prompt_cache_hit_tokens": 0,
                "prompt_cache_miss_tokens": 13
            },
            "system_fingerprint": "fp_4b6881f2c5"
        }
        "#;
        let jd = &mut serde_json::Deserializer::from_str(data);
        let result: Result<CompletionResponse, _> = serde_path_to_error::deserialize(jd);

        match result {
            Ok(response) => match &response.choices.first().unwrap().message {
                Message::Assistant { content, .. } => assert_eq!(
                    content,
                    "Why don’t skeletons fight each other?  \nBecause they don’t have the guts! 😄"
                ),
                _ => panic!("Expected assistant message"),
            },
            Err(err) => {
                panic!("Deserialization error at {}: {}", err.path(), err);
            }
        }
    }

    #[test]
    fn test_serialize_deserialize_tool_call_message() {
        let tool_call_choice_json = r#"
            {
              "finish_reason": "tool_calls",
              "index": 0,
              "logprobs": null,
              "message": {
                "content": "",
                "role": "assistant",
                "tool_calls": [
                  {
                    "function": {
                      "arguments": "{\"x\":2,\"y\":5}",
                      "name": "subtract"
                    },
                    "id": "call_0_2b4a85ee-b04a-40ad-a16b-a405caf6e65b",
                    "index": 0,
                    "type": "function"
                  }
                ]
              }
            }
        "#;

        let choice: Choice = serde_json::from_str(tool_call_choice_json).unwrap();

        let expected_choice: Choice = Choice {
            finish_reason: "tool_calls".to_string(),
            index: 0,
            logprobs: None,
            message: Message::Assistant {
                content: "".to_string(),
                name: None,
                tool_calls: vec![ToolCall {
                    id: "call_0_2b4a85ee-b04a-40ad-a16b-a405caf6e65b".to_string(),
                    function: Function {
                        name: "subtract".to_string(),
                        arguments: serde_json::from_str(r#"{"x":2,"y":5}"#).unwrap(),
                    },
                    index: 0,
                    r#type: ToolType::Function,
                }],
                reasoning_content: None,
            },
        };

        assert_eq!(choice, expected_choice);
    }
}