rig_core/completion/

request.rs

//! Completion request, response, and provider trait definitions.
//!
//! Most applications use [`Prompt`] or [`Chat`] through
//! [`Agent`](crate::agent::Agent). Provider integrations implement
//! [`CompletionModel`] and translate [`CompletionRequest`] into their native HTTP
//! request format.
//!
//! # Low-level request example
//!
//! ```no_run
//! use rig_core::{
//!     client::{CompletionClient, ProviderClient},
//!     completion::{AssistantContent, CompletionModel},
//!     providers::openai,
//! };
//!
//! # async fn run() -> Result<(), Box<dyn std::error::Error>> {
//! let client = openai::Client::from_env()?;
//! let model = client.completion_model(openai::GPT_5_2);
//!
//! let request = model
//!     .completion_request("Who are you?")
//!     .preamble("You are a concise assistant.".to_string())
//!     .temperature(0.5)
//!     .build();
//!
//! let response = model.completion(request).await?;
//! for item in response.choice {
//!     if let AssistantContent::Text(text) = item {
//!         println!("{}", text.text);
//!     }
//! }
//! # Ok(())
//! # }
//! ```

use super::message::{AssistantContent, DocumentMediaType};
use crate::message::ToolChoice;
use crate::streaming::StreamingCompletionResponse;
use crate::tool::server::ToolServerError;
use crate::wasm_compat::{WasmCompatSend, WasmCompatSync};
use crate::{OneOrMany, http_client};
use crate::{
    json_utils,
    message::{Message, UserContent},
    tool::ToolSetError,
};
use serde::de::DeserializeOwned;
use serde::{Deserialize, Serialize};
use std::collections::HashMap;
use std::ops::{Add, AddAssign};
use thiserror::Error;

// Errors
#[derive(Debug, Error)]
pub enum CompletionError {
    /// Http error (e.g.: connection error, timeout, etc.)
    #[error("HttpError: {0}")]
    HttpError(#[from] http_client::Error),

    /// Json error (e.g.: serialization, deserialization)
    #[error("JsonError: {0}")]
    JsonError(#[from] serde_json::Error),

    /// Url error (e.g.: invalid URL)
    #[error("UrlError: {0}")]
    UrlError(#[from] url::ParseError),

    #[cfg(not(target_family = "wasm"))]
    /// Error building the completion request
    #[error("RequestError: {0}")]
    RequestError(#[from] Box<dyn std::error::Error + Send + Sync + 'static>),

    #[cfg(target_family = "wasm")]
    /// Error building the completion request
    #[error("RequestError: {0}")]
    RequestError(#[from] Box<dyn std::error::Error + 'static>),

    /// Error parsing the completion response
    #[error("ResponseError: {0}")]
    ResponseError(String),

    /// Error returned by the completion model provider
    #[error("ProviderError: {0}")]
    ProviderError(String),
}

/// Prompt errors
#[derive(Debug, Error)]
pub enum PromptError {
    /// Something went wrong with the completion
    #[error("CompletionError: {0}")]
    CompletionError(#[from] CompletionError),

    /// There was an error while using a tool
    #[error("ToolCallError: {0}")]
    ToolError(#[from] ToolSetError),

    /// There was an issue while executing a tool on a tool server
    #[error("ToolServerError: {0}")]
    ToolServerError(#[from] Box<ToolServerError>),

    /// The LLM tried to call too many tools during a multi-turn conversation.
    /// To fix this, you may either need to lower the amount of tools your model has access to (and then create other agents to share the tool load)
    /// or increase the amount of turns given in `.multi_turn()`.
    #[error("MaxTurnError: (reached max turn limit: {max_turns})")]
    MaxTurnsError {
        max_turns: usize,
        chat_history: Box<Vec<Message>>,
        prompt: Box<Message>,
    },

    /// A prompting loop was cancelled.
    #[error("PromptCancelled: {reason}")]
    PromptCancelled {
        chat_history: Vec<Message>,
        reason: String,
    },
}

/// Surface [`crate::memory::ConversationMemory`] failures through the existing
/// [`CompletionError::RequestError`] variant so adding memory support does not
/// require a new top-level [`PromptError`] arm in downstream exhaustive matchers.
impl From<crate::memory::MemoryError> for PromptError {
    fn from(err: crate::memory::MemoryError) -> Self {
        Self::CompletionError(CompletionError::RequestError(Box::new(err)))
    }
}

impl PromptError {
    pub(crate) fn prompt_cancelled(
        chat_history: impl IntoIterator<Item = Message>,
        reason: impl Into<String>,
    ) -> Self {
        Self::PromptCancelled {
            chat_history: chat_history.into_iter().collect(),
            reason: reason.into(),
        }
    }
}

/// Errors that can occur when using typed structured output via [`TypedPrompt::prompt_typed`].
#[derive(Debug, Error)]
pub enum StructuredOutputError {
    /// An error occurred during the prompt execution.
    #[error("PromptError: {0}")]
    PromptError(#[from] Box<PromptError>),

    /// Failed to deserialize the model's response into the target type.
    #[error("DeserializationError: {0}")]
    DeserializationError(#[from] serde_json::Error),

    /// The model returned an empty response.
    #[error("EmptyResponse: model returned no content")]
    EmptyResponse,
}

#[derive(Clone, Debug, Deserialize, Serialize)]
pub struct Document {
    /// Stable document identifier included in the serialized context block.
    pub id: String,
    /// Text content passed to the model as retrieval or static context.
    pub text: String,
    /// Additional string metadata rendered before the document text.
    #[serde(flatten)]
    pub additional_props: HashMap<String, String>,
}

impl std::fmt::Display for Document {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        write!(
            f,
            concat!("<file id: {}>\n", "{}\n", "</file>\n"),
            self.id,
            if self.additional_props.is_empty() {
                self.text.clone()
            } else {
                let mut sorted_props = self.additional_props.iter().collect::<Vec<_>>();
                sorted_props.sort_by(|a, b| a.0.cmp(b.0));
                let metadata = sorted_props
                    .iter()
                    .map(|(k, v)| format!("{k}: {v:?}"))
                    .collect::<Vec<_>>()
                    .join(" ");
                format!("<metadata {} />\n{}", metadata, self.text)
            }
        )
    }
}

#[derive(Clone, Debug, Deserialize, Serialize, PartialEq)]
pub struct ToolDefinition {
    /// Tool name exposed to the model. It must match the registered tool name.
    pub name: String,
    /// Human-readable description sent to the model.
    pub description: String,
    /// JSON Schema describing tool arguments.
    pub parameters: serde_json::Value,
}

/// Provider-native tool definition.
///
/// Stored under `additional_params.tools` and forwarded by providers that support
/// provider-managed tools.
#[derive(Clone, Debug, Deserialize, Serialize, PartialEq)]
pub struct ProviderToolDefinition {
    /// Tool type/kind name as expected by the target provider (for example `web_search`).
    #[serde(rename = "type")]
    pub kind: String,
    /// Additional provider-specific configuration for this hosted tool.
    #[serde(flatten, default, skip_serializing_if = "serde_json::Map::is_empty")]
    pub config: serde_json::Map<String, serde_json::Value>,
}

impl ProviderToolDefinition {
    /// Creates a provider-hosted tool definition by type.
    pub fn new(kind: impl Into<String>) -> Self {
        Self {
            kind: kind.into(),
            config: serde_json::Map::new(),
        }
    }

    /// Adds a provider-specific configuration key/value.
    pub fn with_config(mut self, key: impl Into<String>, value: serde_json::Value) -> Self {
        self.config.insert(key.into(), value);
        self
    }
}

// ================================================================
// Implementations
// ================================================================
/// Trait defining a high-level LLM simple prompt interface (i.e.: prompt in, response out).
pub trait Prompt: WasmCompatSend + WasmCompatSync {
    /// Send a simple prompt to the underlying completion model.
    ///
    /// If the completion model's response is a message, then it is returned as a string.
    ///
    /// If the completion model's response is a tool call, then the tool is called and
    /// the result is returned as a string.
    ///
    /// If the tool does not exist, or the tool call fails, then an error is returned.
    fn prompt(
        &self,
        prompt: impl Into<Message> + WasmCompatSend,
    ) -> impl std::future::IntoFuture<Output = Result<String, PromptError>, IntoFuture: WasmCompatSend>;
}

/// Trait defining a high-level LLM chat interface (i.e.: prompt and chat history in, response out).
pub trait Chat: WasmCompatSend + WasmCompatSync {
    /// Send a prompt with optional chat history to the underlying completion model.
    ///
    /// If the completion model's response is a message, then it is returned as a string.
    ///
    /// If the completion model's response is a tool call, then the tool is called and the result
    /// is returned as a string.
    ///
    /// If the tool does not exist, or the tool call fails, then an error is returned.
    ///
    /// The prompt and any assistant or tool messages produced during the turn
    /// are appended to `chat_history`. Callers should pass the current
    /// conversation history and should not push the user prompt themselves
    /// before calling this method.
    fn chat(
        &self,
        prompt: impl Into<Message> + WasmCompatSend,
        chat_history: &mut Vec<Message>,
    ) -> impl std::future::Future<Output = Result<String, PromptError>> + WasmCompatSend;
}

/// Trait defining a high-level typed prompt interface for structured output.
///
/// This trait provides an ergonomic way to get typed responses from an LLM by automatically
/// generating a JSON schema from the target type and deserializing the response.
///
/// # Example
/// ```rust,ignore
/// use rig_core::prelude::*;
/// use schemars::JsonSchema;
/// use serde::Deserialize;
///
/// #[derive(Debug, Deserialize, JsonSchema)]
/// struct WeatherForecast {
///     city: String,
///     temperature_f: f64,
///     conditions: String,
/// }
///
/// let agent = client.agent("gpt-4o").build();
/// let forecast: WeatherForecast = agent
///     .prompt_typed("What's the weather in NYC?")
///     .await?;
/// ```
pub trait TypedPrompt: WasmCompatSend + WasmCompatSync {
    /// The type of the typed prompt request returned by `prompt_typed`.
    type TypedRequest<T>: std::future::IntoFuture<Output = Result<T, StructuredOutputError>>
    where
        T: schemars::JsonSchema + DeserializeOwned + WasmCompatSend + 'static;

    /// Send a prompt and receive a typed structured response.
    ///
    /// The JSON schema for `T` is automatically generated and sent to the provider.
    /// Providers that support native structured outputs will constrain the model's
    /// response to match this schema.
    ///
    /// # Type Parameters
    /// * `T` - The target type to deserialize the response into. Must implement
    ///   `JsonSchema` (for schema generation), `DeserializeOwned` (for deserialization),
    ///   and `WasmCompatSend` (for async compatibility).
    ///
    /// # Example
    /// ```rust,ignore
    /// // Type can be inferred
    /// let forecast: WeatherForecast = agent.prompt_typed("What's the weather?").await?;
    ///
    /// // Or specified explicitly with turbofish
    /// let forecast = agent.prompt_typed::<WeatherForecast>("What's the weather?").await?;
    /// ```
    fn prompt_typed<T>(&self, prompt: impl Into<Message> + WasmCompatSend) -> Self::TypedRequest<T>
    where
        T: schemars::JsonSchema + DeserializeOwned + WasmCompatSend;
}

/// Trait defining a low-level LLM completion interface
pub trait Completion<M: CompletionModel> {
    /// Generates a completion request builder for the given `prompt` and `chat_history`.
    /// This function is meant to be called by the user to further customize the
    /// request at prompt time before sending it.
    ///
    /// ❗IMPORTANT: The type that implements this trait might have already
    /// populated fields in the builder (the exact fields depend on the type).
    /// For fields that have already been set by the model, calling the corresponding
    /// method on the builder will overwrite the value set by the model.
    ///
    /// For example, the request builder returned by [`Agent::completion`](crate::agent::Agent::completion) will already
    /// contain the `preamble` provided when creating the agent.
    fn completion<I, T>(
        &self,
        prompt: impl Into<Message> + WasmCompatSend,
        chat_history: I,
    ) -> impl std::future::Future<Output = Result<CompletionRequestBuilder<M>, CompletionError>>
    + WasmCompatSend
    where
        I: IntoIterator<Item = T> + WasmCompatSend,
        T: Into<Message>;
}

/// General completion response struct that contains the high-level completion choice
/// and the raw response. The completion choice contains one or more assistant content.
#[derive(Debug)]
pub struct CompletionResponse<T> {
    /// The completion choice (represented by one or more assistant message content)
    /// returned by the completion model provider
    pub choice: OneOrMany<AssistantContent>,
    /// Tokens used during prompting and responding
    pub usage: Usage,
    /// The raw response returned by the completion model provider
    pub raw_response: T,
    /// Provider-assigned message ID (e.g. OpenAI Responses API `msg_` ID).
    /// Used to pair reasoning input items with their output items in multi-turn.
    pub message_id: Option<String>,
}

/// A trait for grabbing the token usage of a completion response.
///
/// Primarily designed for streamed completion responses in streamed multi-turn, as otherwise it would be impossible to do.
pub trait GetTokenUsage {
    /// Returns token usage when the response type carries it.
    fn token_usage(&self) -> Option<crate::completion::Usage>;
}

impl GetTokenUsage for () {
    fn token_usage(&self) -> Option<crate::completion::Usage> {
        None
    }
}

impl<T> GetTokenUsage for Option<T>
where
    T: GetTokenUsage,
{
    fn token_usage(&self) -> Option<crate::completion::Usage> {
        if let Some(usage) = self {
            usage.token_usage()
        } else {
            None
        }
    }
}

/// Struct representing the token usage for a completion request.
/// If tokens used are `0`, then the provider failed to supply token usage metrics.
#[derive(Debug, PartialEq, Eq, Clone, Copy, Serialize, Deserialize)]
pub struct Usage {
    /// The number of input ("prompt") tokens used in a given request.
    pub input_tokens: u64,
    /// The number of output ("completion") tokens used in a given request.
    pub output_tokens: u64,
    /// We store this separately as some providers may only report one number
    pub total_tokens: u64,
    /// The number of input tokens read from a provider-managed cache
    pub cached_input_tokens: u64,
    /// The number of input tokens written to a provider-managed cache
    pub cache_creation_input_tokens: u64,
    /// The number of tokens spent on internal reasoning / "thoughts" by reasoning-capable
    /// models (e.g. Gemini thinking, Anthropic extended thinking, OpenAI o-series).
    pub reasoning_tokens: u64,
}

impl Usage {
    /// Creates a new instance of `Usage`.
    pub fn new() -> Self {
        Self {
            input_tokens: 0,
            output_tokens: 0,
            total_tokens: 0,
            cached_input_tokens: 0,
            cache_creation_input_tokens: 0,
            reasoning_tokens: 0,
        }
    }
}

impl Default for Usage {
    fn default() -> Self {
        Self::new()
    }
}

impl Add for Usage {
    type Output = Self;

    fn add(self, other: Self) -> Self::Output {
        Self {
            input_tokens: self.input_tokens + other.input_tokens,
            output_tokens: self.output_tokens + other.output_tokens,
            total_tokens: self.total_tokens + other.total_tokens,
            cached_input_tokens: self.cached_input_tokens + other.cached_input_tokens,
            cache_creation_input_tokens: self.cache_creation_input_tokens
                + other.cache_creation_input_tokens,
            reasoning_tokens: self.reasoning_tokens + other.reasoning_tokens,
        }
    }
}

impl AddAssign for Usage {
    fn add_assign(&mut self, other: Self) {
        self.input_tokens += other.input_tokens;
        self.output_tokens += other.output_tokens;
        self.total_tokens += other.total_tokens;
        self.cached_input_tokens += other.cached_input_tokens;
        self.cache_creation_input_tokens += other.cache_creation_input_tokens;
        self.reasoning_tokens += other.reasoning_tokens;
    }
}

/// Trait defining a completion model that can be used to generate completion responses.
/// This trait is meant to be implemented by the user to define a custom completion model,
/// either from a third party provider (e.g.: OpenAI) or a local model.
pub trait CompletionModel: Clone + WasmCompatSend + WasmCompatSync {
    /// The raw response type returned by the underlying completion model.
    type Response: WasmCompatSend + WasmCompatSync + Serialize + DeserializeOwned;
    /// The raw response type returned by the underlying completion model when streaming.
    type StreamingResponse: Clone
        + Unpin
        + WasmCompatSend
        + WasmCompatSync
        + Serialize
        + DeserializeOwned
        + GetTokenUsage;

    /// Provider client type used to construct this model.
    type Client;

    /// Construct a model handle from a provider client and model identifier.
    fn make(client: &Self::Client, model: impl Into<String>) -> Self;

    /// Generates a completion response for the given completion request.
    fn completion(
        &self,
        request: CompletionRequest,
    ) -> impl std::future::Future<
        Output = Result<CompletionResponse<Self::Response>, CompletionError>,
    > + WasmCompatSend;

    fn stream(
        &self,
        request: CompletionRequest,
    ) -> impl std::future::Future<
        Output = Result<StreamingCompletionResponse<Self::StreamingResponse>, CompletionError>,
    > + WasmCompatSend;

    /// Generates a completion request builder for the given `prompt`.
    fn completion_request(&self, prompt: impl Into<Message>) -> CompletionRequestBuilder<Self> {
        CompletionRequestBuilder::new(self.clone(), prompt)
    }
}

/// Struct representing a general completion request that can be sent to a completion model provider.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct CompletionRequest {
    /// Optional model override for this request.
    pub model: Option<String>,
    /// Legacy preamble field preserved for backwards compatibility.
    ///
    /// New code should prefer a leading [`Message::System`]
    /// in `chat_history` as the canonical representation of system instructions.
    pub preamble: Option<String>,
    /// The chat history to be sent to the completion model provider.
    /// The very last message will always be the prompt (hence why there is *always* one)
    pub chat_history: OneOrMany<Message>,
    /// The documents to be sent to the completion model provider
    pub documents: Vec<Document>,
    /// The tools to be sent to the completion model provider
    pub tools: Vec<ToolDefinition>,
    /// The temperature to be sent to the completion model provider
    pub temperature: Option<f64>,
    /// The max tokens to be sent to the completion model provider
    pub max_tokens: Option<u64>,
    /// Whether tools are required to be used by the model provider or not before providing a response.
    pub tool_choice: Option<ToolChoice>,
    /// Additional provider-specific parameters to be sent to the completion model provider
    pub additional_params: Option<serde_json::Value>,
    /// Optional JSON Schema for structured output. When set, providers that support
    /// native structured outputs will constrain the model's response to match this schema.
    pub output_schema: Option<schemars::Schema>,
}

impl CompletionRequest {
    /// Extracts a name from the output schema's `"title"` field, falling back to `"response_schema"`.
    /// Useful for providers that require a name alongside the JSON Schema (e.g., OpenAI).
    pub fn output_schema_name(&self) -> Option<String> {
        self.output_schema.as_ref().map(|schema| {
            schema
                .as_object()
                .and_then(|o| o.get("title"))
                .and_then(|v| v.as_str())
                .unwrap_or("response_schema")
                .to_string()
        })
    }

    /// Returns documents normalized into a message (if any).
    /// Most providers do not accept documents directly as input, so it needs to convert into a
    /// `Message` so that it can be incorporated into `chat_history`.
    pub fn normalized_documents(&self) -> Option<Message> {
        if self.documents.is_empty() {
            return None;
        }

        // Most providers will convert documents into a text unless it can handle document messages.
        // We use `UserContent::document` for those who handle it directly!
        let messages = self
            .documents
            .iter()
            .map(|doc| {
                UserContent::document(
                    doc.to_string(),
                    // In the future, we can customize `Document` to pass these extra types through.
                    // Most providers ditch these but they might want to use them.
                    Some(DocumentMediaType::TXT),
                )
            })
            .collect::<Vec<_>>();

        OneOrMany::from_iter_optional(messages).map(|content| Message::User { content })
    }

    /// Adds a provider-hosted tool by storing it in `additional_params.tools`.
    pub fn with_provider_tool(mut self, tool: ProviderToolDefinition) -> Self {
        self.additional_params =
            merge_provider_tools_into_additional_params(self.additional_params, vec![tool]);
        self
    }

    /// Adds provider-hosted tools by storing them in `additional_params.tools`.
    pub fn with_provider_tools(mut self, tools: Vec<ProviderToolDefinition>) -> Self {
        self.additional_params =
            merge_provider_tools_into_additional_params(self.additional_params, tools);
        self
    }
}

fn merge_provider_tools_into_additional_params(
    additional_params: Option<serde_json::Value>,
    provider_tools: Vec<ProviderToolDefinition>,
) -> Option<serde_json::Value> {
    if provider_tools.is_empty() {
        return additional_params;
    }

    let mut provider_tools_json = provider_tools
        .into_iter()
        .map(|ProviderToolDefinition { kind, mut config }| {
            // Force the provider tool type from the strongly-typed field.
            config.insert("type".to_string(), serde_json::Value::String(kind));
            serde_json::Value::Object(config)
        })
        .collect::<Vec<_>>();

    let mut params_map = match additional_params {
        Some(serde_json::Value::Object(map)) => map,
        Some(serde_json::Value::Bool(stream)) => {
            let mut map = serde_json::Map::new();
            map.insert("stream".to_string(), serde_json::Value::Bool(stream));
            map
        }
        _ => serde_json::Map::new(),
    };

    let mut merged_tools = match params_map.remove("tools") {
        Some(serde_json::Value::Array(existing)) => existing,
        _ => Vec::new(),
    };
    merged_tools.append(&mut provider_tools_json);
    params_map.insert("tools".to_string(), serde_json::Value::Array(merged_tools));
    Some(serde_json::Value::Object(params_map))
}

/// Builder struct for constructing a completion request.
///
/// Example usage:
/// ```no_run
/// use rig_core::{
///     client::CompletionClient,
///     providers::openai::{Client, self},
///     completion::{CompletionModel, CompletionRequestBuilder},
/// };
///
/// # async fn run() -> Result<(), Box<dyn std::error::Error>> {
/// let openai = Client::new("your-openai-api-key")?;
/// let model = openai.completion_model(openai::GPT_5_2);
///
/// // Create the completion request and execute it separately
/// let request = CompletionRequestBuilder::new(model.clone(), "Who are you?".to_string())
///     .preamble("You are Marvin from the Hitchhiker's Guide to the Galaxy.".to_string())
///     .temperature(0.5)
///     .build();
///
/// let response = model.completion(request).await?;
/// # Ok(())
/// # }
/// ```
///
/// Alternatively, you can execute the completion request directly from the builder:
/// ```no_run
/// use rig_core::{
///     client::CompletionClient,
///     providers::openai::{Client, self},
///     completion::CompletionRequestBuilder,
/// };
///
/// # async fn run() -> Result<(), Box<dyn std::error::Error>> {
/// let openai = Client::new("your-openai-api-key")?;
/// let model = openai.completion_model(openai::GPT_5_2);
///
/// // Create the completion request and execute it directly
/// let response = CompletionRequestBuilder::new(model, "Who are you?".to_string())
///     .preamble("You are Marvin from the Hitchhiker's Guide to the Galaxy.".to_string())
///     .temperature(0.5)
///     .send()
///     .await?;
/// # Ok(())
/// # }
/// ```
///
/// Note: It is usually unnecessary to create a completion request builder directly.
/// Instead, use the [CompletionModel::completion_request] method.
pub struct CompletionRequestBuilder<M: CompletionModel> {
    model: M,
    prompt: Message,
    request_model: Option<String>,
    preamble: Option<String>,
    chat_history: Vec<Message>,
    documents: Vec<Document>,
    tools: Vec<ToolDefinition>,
    provider_tools: Vec<ProviderToolDefinition>,
    temperature: Option<f64>,
    max_tokens: Option<u64>,
    tool_choice: Option<ToolChoice>,
    additional_params: Option<serde_json::Value>,
    output_schema: Option<schemars::Schema>,
}

impl<M: CompletionModel> CompletionRequestBuilder<M> {
    pub fn new(model: M, prompt: impl Into<Message>) -> Self {
        Self {
            model,
            prompt: prompt.into(),
            request_model: None,
            preamble: None,
            chat_history: Vec::new(),
            documents: Vec::new(),
            tools: Vec::new(),
            provider_tools: Vec::new(),
            temperature: None,
            max_tokens: None,
            tool_choice: None,
            additional_params: None,
            output_schema: None,
        }
    }

    /// Sets the preamble for the completion request.
    pub fn preamble(mut self, preamble: String) -> Self {
        // Legacy public API: funnel preamble into canonical system messages at build-time.
        self.preamble = Some(preamble);
        self
    }

    /// Overrides the model used for this request.
    pub fn model(mut self, model: impl Into<String>) -> Self {
        self.request_model = Some(model.into());
        self
    }

    /// Overrides the model used for this request.
    pub fn model_opt(mut self, model: Option<String>) -> Self {
        self.request_model = model;
        self
    }

    pub fn without_preamble(mut self) -> Self {
        self.preamble = None;
        self
    }

    /// Adds a message to the chat history for the completion request.
    pub fn message(mut self, message: Message) -> Self {
        self.chat_history.push(message);

        self
    }

    /// Adds a list of messages to the chat history for the completion request.
    pub fn messages(mut self, messages: impl IntoIterator<Item = Message>) -> Self {
        self.chat_history.extend(messages);

        self
    }

    /// Adds a document to the completion request.
    pub fn document(mut self, document: Document) -> Self {
        self.documents.push(document);
        self
    }

    /// Adds a list of documents to the completion request.
    pub fn documents(self, documents: impl IntoIterator<Item = Document>) -> Self {
        documents
            .into_iter()
            .fold(self, |builder, doc| builder.document(doc))
    }

    /// Adds a tool to the completion request.
    pub fn tool(mut self, tool: ToolDefinition) -> Self {
        self.tools.push(tool);
        self
    }

    /// Adds a list of tools to the completion request.
    pub fn tools(self, tools: Vec<ToolDefinition>) -> Self {
        tools
            .into_iter()
            .fold(self, |builder, tool| builder.tool(tool))
    }

    /// Adds a provider-hosted tool to the completion request.
    pub fn provider_tool(mut self, tool: ProviderToolDefinition) -> Self {
        self.provider_tools.push(tool);
        self
    }

    /// Adds provider-hosted tools to the completion request.
    pub fn provider_tools(self, tools: Vec<ProviderToolDefinition>) -> Self {
        tools
            .into_iter()
            .fold(self, |builder, tool| builder.provider_tool(tool))
    }

    /// Adds additional parameters to the completion request.
    /// This can be used to set additional provider-specific parameters. For example,
    /// Cohere's completion models accept a `connectors` parameter that can be used to
    /// specify the data connectors used by Cohere when executing the completion
    /// (see `examples/cohere_connectors.rs`).
    pub fn additional_params(mut self, additional_params: serde_json::Value) -> Self {
        match self.additional_params {
            Some(params) => {
                self.additional_params = Some(json_utils::merge(params, additional_params));
            }
            None => {
                self.additional_params = Some(additional_params);
            }
        }
        self
    }

    /// Sets the additional parameters for the completion request.
    /// This can be used to set additional provider-specific parameters. For example,
    /// Cohere's completion models accept a `connectors` parameter that can be used to
    /// specify the data connectors used by Cohere when executing the completion
    /// (see `examples/cohere_connectors.rs`).
    pub fn additional_params_opt(mut self, additional_params: Option<serde_json::Value>) -> Self {
        self.additional_params = additional_params;
        self
    }

    /// Sets the temperature for the completion request.
    pub fn temperature(mut self, temperature: f64) -> Self {
        self.temperature = Some(temperature);
        self
    }

    /// Sets the temperature for the completion request.
    pub fn temperature_opt(mut self, temperature: Option<f64>) -> Self {
        self.temperature = temperature;
        self
    }

    /// Sets the max tokens for the completion request.
    /// Note: This is required if using Anthropic
    pub fn max_tokens(mut self, max_tokens: u64) -> Self {
        self.max_tokens = Some(max_tokens);
        self
    }

    /// Sets the max tokens for the completion request.
    /// Note: This is required if using Anthropic
    pub fn max_tokens_opt(mut self, max_tokens: Option<u64>) -> Self {
        self.max_tokens = max_tokens;
        self
    }

    /// Sets the thing.
    pub fn tool_choice(mut self, tool_choice: ToolChoice) -> Self {
        self.tool_choice = Some(tool_choice);
        self
    }

    /// Sets the output schema for structured output. When set, providers that support
    /// native structured outputs will constrain the model's response to match this schema.
    /// NOTE: For direct type conversion, you may want to use `Agent::prompt_typed()` - using this method
    /// with `Agent::prompt()` will still output a String at the end, it'll just be compatible with whatever
    /// type you want to use here. This method is primarily an escape hatch for agents being used as tools
    /// to still be able to leverage structured outputs.
    pub fn output_schema(mut self, schema: schemars::Schema) -> Self {
        self.output_schema = Some(schema);
        self
    }

    /// Sets the output schema for structured output from an optional value.
    /// NOTE: For direct type conversion, you may want to use `Agent::prompt_typed()` - using this method
    /// with `Agent::prompt()` will still output a String at the end, it'll just be compatible with whatever
    /// type you want to use here. This method is primarily an escape hatch for agents being used as tools
    /// to still be able to leverage structured outputs.
    pub fn output_schema_opt(mut self, schema: Option<schemars::Schema>) -> Self {
        self.output_schema = schema;
        self
    }

    /// Builds the completion request.
    pub fn build(self) -> CompletionRequest {
        // Build the final message list, prepending preamble if present
        let mut chat_history = self.chat_history;
        let prompt = self.prompt;
        if let Some(preamble) = self.preamble {
            chat_history.insert(0, Message::system(preamble));
        }
        chat_history.push(prompt.clone());

        let chat_history =
            OneOrMany::from_iter_optional(chat_history).unwrap_or_else(|| OneOrMany::one(prompt));
        let additional_params = merge_provider_tools_into_additional_params(
            self.additional_params,
            self.provider_tools,
        );

        CompletionRequest {
            model: self.request_model,
            preamble: None,
            chat_history,
            documents: self.documents,
            tools: self.tools,
            temperature: self.temperature,
            max_tokens: self.max_tokens,
            tool_choice: self.tool_choice,
            additional_params,
            output_schema: self.output_schema,
        }
    }

    /// Sends the completion request to the completion model provider and returns the completion response.
    pub async fn send(self) -> Result<CompletionResponse<M::Response>, CompletionError> {
        let model = self.model.clone();
        model.completion(self.build()).await
    }

    /// Stream the completion request
    pub async fn stream<'a>(
        self,
    ) -> Result<StreamingCompletionResponse<M::StreamingResponse>, CompletionError>
    where
        <M as CompletionModel>::StreamingResponse: 'a,
        Self: 'a,
    {
        let model = self.model.clone();
        model.stream(self.build()).await
    }
}

#[cfg(test)]
mod tests {

    use super::*;
    use crate::test_utils::MockCompletionModel;

    #[test]
    fn test_document_display_without_metadata() {
        let doc = Document {
            id: "123".to_string(),
            text: "This is a test document.".to_string(),
            additional_props: HashMap::new(),
        };

        let expected = "<file id: 123>\nThis is a test document.\n</file>\n";
        assert_eq!(format!("{doc}"), expected);
    }

    #[test]
    fn test_document_display_with_metadata() {
        let mut additional_props = HashMap::new();
        additional_props.insert("author".to_string(), "John Doe".to_string());
        additional_props.insert("length".to_string(), "42".to_string());

        let doc = Document {
            id: "123".to_string(),
            text: "This is a test document.".to_string(),
            additional_props,
        };

        let expected = concat!(
            "<file id: 123>\n",
            "<metadata author: \"John Doe\" length: \"42\" />\n",
            "This is a test document.\n",
            "</file>\n"
        );
        assert_eq!(format!("{doc}"), expected);
    }

    #[test]
    fn test_normalize_documents_with_documents() {
        let doc1 = Document {
            id: "doc1".to_string(),
            text: "Document 1 text.".to_string(),
            additional_props: HashMap::new(),
        };

        let doc2 = Document {
            id: "doc2".to_string(),
            text: "Document 2 text.".to_string(),
            additional_props: HashMap::new(),
        };

        let request = CompletionRequest {
            model: None,
            preamble: None,
            chat_history: OneOrMany::one("What is the capital of France?".into()),
            documents: vec![doc1, doc2],
            tools: Vec::new(),
            temperature: None,
            max_tokens: None,
            tool_choice: None,
            additional_params: None,
            output_schema: None,
        };

        let expected = Message::User {
            content: OneOrMany::many(vec![
                UserContent::document(
                    "<file id: doc1>\nDocument 1 text.\n</file>\n".to_string(),
                    Some(DocumentMediaType::TXT),
                ),
                UserContent::document(
                    "<file id: doc2>\nDocument 2 text.\n</file>\n".to_string(),
                    Some(DocumentMediaType::TXT),
                ),
            ])
            .expect("There will be at least one document"),
        };

        assert_eq!(request.normalized_documents(), Some(expected));
    }

    #[test]
    fn test_normalize_documents_without_documents() {
        let request = CompletionRequest {
            model: None,
            preamble: None,
            chat_history: OneOrMany::one("What is the capital of France?".into()),
            documents: Vec::new(),
            tools: Vec::new(),
            temperature: None,
            max_tokens: None,
            tool_choice: None,
            additional_params: None,
            output_schema: None,
        };

        assert_eq!(request.normalized_documents(), None);
    }

    #[test]
    fn preamble_builder_funnels_to_system_message() {
        let request =
            CompletionRequestBuilder::new(MockCompletionModel::default(), Message::user("Prompt"))
                .preamble("System prompt".to_string())
                .message(Message::user("History"))
                .build();

        assert_eq!(request.preamble, None);

        let history = request.chat_history.into_iter().collect::<Vec<_>>();
        assert_eq!(history.len(), 3);
        assert!(matches!(
            &history[0],
            Message::System { content } if content == "System prompt"
        ));
        assert!(matches!(&history[1], Message::User { .. }));
        assert!(matches!(&history[2], Message::User { .. }));
    }

    #[test]
    fn without_preamble_removes_legacy_preamble_injection() {
        let request =
            CompletionRequestBuilder::new(MockCompletionModel::default(), Message::user("Prompt"))
                .preamble("System prompt".to_string())
                .without_preamble()
                .build();

        assert_eq!(request.preamble, None);
        let history = request.chat_history.into_iter().collect::<Vec<_>>();
        assert_eq!(history.len(), 1);
        assert!(matches!(&history[0], Message::User { .. }));
    }
}