litellm-rs 0.1.1

//! Cohere provider implementation
//!
//! This module provides Cohere API integration.

use super::{BaseProvider, ModelPricing, Provider, ProviderError, ProviderType};
use crate::config::ProviderConfig;
use crate::core::models::{RequestContext, openai::*};
use crate::utils::error::Result;
use async_trait::async_trait;
use serde_json::json;
use std::collections::HashMap;
use tracing::{debug, info};

/// Cohere provider implementation
#[derive(Debug, Clone)]
pub struct CohereProvider {
    /// Base provider functionality
    base: BaseProvider,
    /// Model pricing cache
    pricing_cache: HashMap<String, ModelPricing>,
}

impl CohereProvider {
    /// Create a new Cohere provider
    pub async fn new(config: &ProviderConfig) -> Result<Self> {
        let base = BaseProvider::new(config)?;

        // Set default base URL if not provided
        let base_url = config
            .base_url
            .clone()
            .unwrap_or_else(|| "https://api.cohere.ai/v1".to_string());

        let provider = Self {
            base: BaseProvider { base_url, ..base },
            pricing_cache: Self::initialize_pricing_cache(),
        };

        info!("Cohere provider '{}' initialized successfully", config.name);
        Ok(provider)
    }

    /// Initialize pricing cache with known Cohere model prices
    fn initialize_pricing_cache() -> HashMap<String, ModelPricing> {
        let mut cache = HashMap::new();

        // Command models
        cache.insert(
            "command".to_string(),
            ModelPricing {
                model: "command".to_string(),
                input_cost_per_1k: 0.0015,
                output_cost_per_1k: 0.002,
                currency: "USD".to_string(),
                updated_at: chrono::Utc::now(),
            },
        );

        cache.insert(
            "command-light".to_string(),
            ModelPricing {
                model: "command-light".to_string(),
                input_cost_per_1k: 0.0003,
                output_cost_per_1k: 0.0006,
                currency: "USD".to_string(),
                updated_at: chrono::Utc::now(),
            },
        );

        // Embedding models
        cache.insert(
            "embed-english-v3.0".to_string(),
            ModelPricing {
                model: "embed-english-v3.0".to_string(),
                input_cost_per_1k: 0.0001,
                output_cost_per_1k: 0.0,
                currency: "USD".to_string(),
                updated_at: chrono::Utc::now(),
            },
        );

        cache
    }

    /// Convert OpenAI messages to Cohere format
    fn convert_messages_to_cohere(&self, messages: &[ChatMessage]) -> (Option<String>, String) {
        let mut system_message = None;
        let mut conversation_parts = Vec::new();

        for message in messages {
            match message.role {
                MessageRole::System => {
                    if let Some(MessageContent::Text(text)) = &message.content {
                        system_message = Some(text.clone());
                    }
                }
                MessageRole::User => {
                    if let Some(content) = &message.content {
                        conversation_parts
                            .push(format!("User: {}", self.extract_text_content(content)));
                    }
                }
                MessageRole::Assistant => {
                    if let Some(content) = &message.content {
                        conversation_parts
                            .push(format!("Assistant: {}", self.extract_text_content(content)));
                    }
                }
                _ => {}
            }
        }

        let conversation = conversation_parts.join("\n");
        (system_message, conversation)
    }

    /// Extract text content from MessageContent
    fn extract_text_content(&self, content: &MessageContent) -> String {
        match content {
            MessageContent::Text(text) => text.clone(),
            MessageContent::Parts(parts) => parts
                .iter()
                .filter_map(|part| match part {
                    ContentPart::Text { text } => Some(text.clone()),
                    _ => None,
                })
                .collect::<Vec<String>>()
                .join("\n"),
        }
    }

    /// Convert Cohere response to OpenAI format
    fn convert_cohere_response_to_openai(
        &self,
        cohere_response: serde_json::Value,
        model: &str,
    ) -> Result<ChatCompletionResponse> {
        let content = cohere_response
            .get("generations")
            .and_then(|g| g.as_array())
            .and_then(|arr| arr.first())
            .and_then(|generation| generation.get("text"))
            .and_then(|text| text.as_str())
            .unwrap_or("")
            .to_string();

        let usage = Usage {
            prompt_tokens: 0, // Cohere doesn't provide detailed token counts
            completion_tokens: 0,
            total_tokens: 0,
            prompt_tokens_details: None,
            completion_tokens_details: None,
        };

        Ok(ChatCompletionResponse {
            id: format!("chatcmpl-cohere-{}", uuid::Uuid::new_v4()),
            object: "chat.completion".to_string(),
            created: chrono::Utc::now().timestamp() as u64,
            model: model.to_string(),
            choices: vec![ChatChoice {
                index: 0,
                message: ChatMessage {
                    role: MessageRole::Assistant,
                    content: Some(MessageContent::Text(content)),
                    name: None,
                    function_call: None,
                    tool_calls: None,
                    tool_call_id: None,
                    audio: None,
                },
                finish_reason: Some("stop".to_string()),
                logprobs: None,
            }],
            usage: Some(usage),
            system_fingerprint: None,
        })
    }

    /// Create request headers for Cohere API
    fn create_headers(&self) -> reqwest::header::HeaderMap {
        let mut headers = reqwest::header::HeaderMap::new();

        headers.insert(
            reqwest::header::AUTHORIZATION,
            format!("Bearer {}", self.base.api_key).parse().unwrap(),
        );

        headers.insert(
            reqwest::header::CONTENT_TYPE,
            "application/json".parse().unwrap(),
        );

        headers
    }
}

#[async_trait]
impl Provider for CohereProvider {
    fn name(&self) -> &str {
        &self.base.name
    }

    fn provider_type(&self) -> ProviderType {
        ProviderType::Cohere
    }

    async fn supports_model(&self, model: &str) -> bool {
        self.base.is_model_supported(model)
            || model.starts_with("command")
            || model.starts_with("embed-")
    }

    async fn supports_images(&self) -> bool {
        false // Cohere doesn't support images
    }

    async fn supports_embeddings(&self) -> bool {
        true
    }

    async fn supports_streaming(&self) -> bool {
        true
    }

    async fn list_models(&self) -> Result<Vec<Model>> {
        // Return known models
        let known_models = vec![
            "command",
            "command-light",
            "command-nightly",
            "embed-english-v3.0",
            "embed-multilingual-v3.0",
        ];

        let models = known_models
            .into_iter()
            .map(|model| Model {
                id: model.to_string(),
                object: "model".to_string(),
                created: chrono::Utc::now().timestamp() as u64,
                owned_by: "cohere".to_string(),
            })
            .collect();

        Ok(models)
    }

    async fn health_check(&self) -> Result<()> {
        debug!("Performing Cohere health check");
        // TODO: Implement Cohere-specific health check
        Ok(())
    }

    async fn chat_completion(
        &self,
        request: ChatCompletionRequest,
        _context: RequestContext,
    ) -> Result<ChatCompletionResponse> {
        debug!("Cohere chat completion for model: {}", request.model);

        let (system_message, _conversation) = self.convert_messages_to_cohere(&request.messages);

        // Get the last user message as the prompt
        let prompt = request
            .messages
            .iter()
            .filter(|msg| msg.role == MessageRole::User)
            .next_back()
            .and_then(|msg| msg.content.as_ref().map(|c| self.extract_text_content(c)))
            .unwrap_or_default();

        let mut body = json!({
            "model": request.model,
            "prompt": prompt,
            "max_tokens": request.max_tokens.unwrap_or(4096),
        });

        if let Some(system) = system_message {
            body["preamble"] = json!(system);
        }

        if let Some(temperature) = request.temperature {
            body["temperature"] = json!(temperature);
        }

        if let Some(top_p) = request.top_p {
            body["p"] = json!(top_p);
        }

        if let Some(stop) = request.stop {
            body["stop_sequences"] = json!(stop);
        }

        let url = format!("{}/generate", self.base.base_url);
        let response = self
            .base
            .client
            .post(&url)
            .headers(self.create_headers())
            .json(&body)
            .send()
            .await
            .map_err(|e| ProviderError::Network(e.to_string()))?;

        if !response.status().is_success() {
            let status = response.status();
            let error_text = response.text().await.unwrap_or_default();

            return Err(match status.as_u16() {
                401 => ProviderError::Authentication(error_text),
                429 => ProviderError::RateLimit(error_text),
                400 => ProviderError::InvalidRequest(error_text),
                _ => ProviderError::Unknown(format!("HTTP {}: {}", status, error_text)),
            }
            .into());
        }

        let cohere_response: serde_json::Value = self.base.parse_json_response(response).await?;
        self.convert_cohere_response_to_openai(cohere_response, &request.model)
    }

    async fn completion(
        &self,
        request: CompletionRequest,
        _context: RequestContext,
    ) -> Result<CompletionResponse> {
        debug!("Cohere completion for model: {}", request.model);

        let body = json!({
            "model": request.model,
            "prompt": request.prompt,
            "max_tokens": request.max_tokens,
            "temperature": request.temperature,
            "p": request.top_p,
            "stop_sequences": request.stop
        });

        let url = format!("{}/generate", self.base.base_url);
        let response = self
            .base
            .client
            .post(&url)
            .headers(self.create_headers())
            .json(&body)
            .send()
            .await
            .map_err(|e| ProviderError::Network(e.to_string()))?;

        let cohere_response: serde_json::Value = self.base.parse_json_response(response).await?;

        // Convert to OpenAI completion format
        let text = cohere_response
            .get("generations")
            .and_then(|g| g.as_array())
            .and_then(|arr| arr.first())
            .and_then(|generation| generation.get("text"))
            .and_then(|text| text.as_str())
            .unwrap_or("")
            .to_string();

        Ok(CompletionResponse {
            id: format!("cmpl-cohere-{}", uuid::Uuid::new_v4()),
            object: "text_completion".to_string(),
            created: chrono::Utc::now().timestamp() as u64,
            model: request.model,
            choices: vec![CompletionChoice {
                text,
                index: 0,
                logprobs: None,
                finish_reason: Some("stop".to_string()),
            }],
            usage: Some(Usage {
                prompt_tokens: 0,
                completion_tokens: 0,
                total_tokens: 0,
                prompt_tokens_details: None,
                completion_tokens_details: None,
            }),
        })
    }

    async fn embedding(
        &self,
        request: EmbeddingRequest,
        _context: RequestContext,
    ) -> Result<EmbeddingResponse> {
        debug!("Cohere embedding for model: {}", request.model);

        let body = json!({
            "model": request.model,
            "texts": request.input
        });

        let url = format!("{}/embed", self.base.base_url);
        let response = self
            .base
            .client
            .post(&url)
            .headers(self.create_headers())
            .json(&body)
            .send()
            .await
            .map_err(|e| ProviderError::Network(e.to_string()))?;

        let cohere_response: serde_json::Value = self.base.parse_json_response(response).await?;

        // Convert to OpenAI embedding format
        let embeddings = cohere_response
            .get("embeddings")
            .and_then(|e| e.as_array())
            .unwrap_or(&vec![])
            .iter()
            .enumerate()
            .map(|(index, embedding)| {
                let embedding_vec = embedding
                    .as_array()
                    .unwrap_or(&vec![])
                    .iter()
                    .filter_map(|v| v.as_f64())
                    .collect();

                EmbeddingObject {
                    object: "embedding".to_string(),
                    embedding: embedding_vec,
                    index: index as u32,
                }
            })
            .collect();

        Ok(EmbeddingResponse {
            object: "list".to_string(),
            data: embeddings,
            model: request.model,
            usage: EmbeddingUsage {
                prompt_tokens: 0,
                total_tokens: 0,
            },
        })
    }

    async fn image_generation(
        &self,
        _request: ImageGenerationRequest,
        _context: RequestContext,
    ) -> Result<ImageGenerationResponse> {
        Err(
            ProviderError::InvalidRequest("Cohere does not support image generation".to_string())
                .into(),
        )
    }

    async fn get_model_pricing(&self, model: &str) -> Result<ModelPricing> {
        if let Some(pricing) = self.pricing_cache.get(model) {
            Ok(pricing.clone())
        } else {
            // Return default pricing for unknown models
            Ok(ModelPricing {
                model: model.to_string(),
                input_cost_per_1k: 0.0015, // Default rate
                output_cost_per_1k: 0.002,
                currency: "USD".to_string(),
                updated_at: chrono::Utc::now(),
            })
        }
    }

    async fn calculate_cost(
        &self,
        model: &str,
        input_tokens: u32,
        output_tokens: u32,
    ) -> Result<f64> {
        let pricing = self.get_model_pricing(model).await?;

        let input_cost = (input_tokens as f64 / 1000.0) * pricing.input_cost_per_1k;
        let output_cost = (output_tokens as f64 / 1000.0) * pricing.output_cost_per_1k;

        Ok(input_cost + output_cost)
    }
}