litellm-rs 0.1.1

//! AI21 Labs provider implementation
//!
//! This module provides AI21 Labs API integration for Jurassic models.

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};

/// AI21 Labs provider implementation
#[derive(Debug, Clone)]
pub struct AI21Provider {
    /// Base provider functionality
    base: BaseProvider,
    /// Model pricing cache
    pricing_cache: HashMap<String, ModelPricing>,
}

impl AI21Provider {
    /// Create a new AI21 provider
    pub async fn new(config: &ProviderConfig) -> Result<Self> {
        let base = BaseProvider::new(config)?;

        let base_url = config
            .base_url
            .clone()
            .unwrap_or_else(|| "https://api.ai21.com".to_string());

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

        info!("AI21 provider '{}' initialized successfully", config.name);
        Ok(provider)
    }

    /// Initialize pricing cache with AI21 model prices
    fn initialize_pricing_cache() -> HashMap<String, ModelPricing> {
        let mut cache = HashMap::new();

        // Jurassic models
        cache.insert(
            "j2-ultra".to_string(),
            ModelPricing {
                model: "j2-ultra".to_string(),
                input_cost_per_1k: 0.015,
                output_cost_per_1k: 0.015,
                currency: "USD".to_string(),
                updated_at: chrono::Utc::now(),
            },
        );

        cache.insert(
            "j2-mid".to_string(),
            ModelPricing {
                model: "j2-mid".to_string(),
                input_cost_per_1k: 0.01,
                output_cost_per_1k: 0.01,
                currency: "USD".to_string(),
                updated_at: chrono::Utc::now(),
            },
        );

        cache.insert(
            "j2-light".to_string(),
            ModelPricing {
                model: "j2-light".to_string(),
                input_cost_per_1k: 0.003,
                output_cost_per_1k: 0.003,
                currency: "USD".to_string(),
                updated_at: chrono::Utc::now(),
            },
        );

        // Jamba models
        cache.insert(
            "jamba-instruct".to_string(),
            ModelPricing {
                model: "jamba-instruct".to_string(),
                input_cost_per_1k: 0.0005,
                output_cost_per_1k: 0.0007,
                currency: "USD".to_string(),
                updated_at: chrono::Utc::now(),
            },
        );

        cache
    }

    /// Convert OpenAI messages to AI21 format
    fn convert_messages_to_ai21(&self, messages: &[ChatMessage]) -> String {
        // AI21 typically expects a single prompt string
        messages
            .iter()
            .map(|msg| {
                let role = match msg.role {
                    MessageRole::System => "System",
                    MessageRole::User => "Human",
                    MessageRole::Assistant => "Assistant",
                    MessageRole::Tool => "Tool",
                    MessageRole::Function => "Function",
                };

                let content = match &msg.content {
                    Some(MessageContent::Text(text)) => text.clone(),
                    Some(MessageContent::Parts(parts)) => parts
                        .iter()
                        .filter_map(|part| match part {
                            ContentPart::Text { text } => Some(text.clone()),
                            _ => None,
                        })
                        .collect::<Vec<String>>()
                        .join(" "),
                    None => String::new(),
                };

                format!("{}: {}", role, content)
            })
            .collect::<Vec<String>>()
            .join("\n\n")
    }

    /// Convert AI21 response to OpenAI format
    fn convert_ai21_response_to_openai(
        &self,
        ai21_response: serde_json::Value,
        model: &str,
    ) -> Result<ChatCompletionResponse> {
        let completions = ai21_response
            .get("completions")
            .and_then(|c| c.as_array())
            .ok_or_else(|| ProviderError::Parsing("No completions in response".to_string()))?;

        let choices: Vec<ChatCompletionChoice> = completions
            .iter()
            .enumerate()
            .map(|(index, completion)| {
                let text = completion
                    .get("data")
                    .and_then(|d| d.get("text"))
                    .and_then(|t| t.as_str())
                    .unwrap_or("")
                    .to_string();

                let finish_reason = completion
                    .get("finishReason")
                    .and_then(|fr| fr.get("reason"))
                    .and_then(|r| r.as_str())
                    .map(|r| match r {
                        "stop" => "stop",
                        "length" => "length",
                        _ => "stop",
                    })
                    .unwrap_or("stop")
                    .to_string();

                ChatCompletionChoice {
                    index: index as u32,
                    message: ChatMessage {
                        role: MessageRole::Assistant,
                        content: Some(MessageContent::Text(text)),
                        name: None,
                        function_call: None,
                        tool_calls: None,
                        tool_call_id: None,
                        audio: None,
                    },
                    finish_reason: Some(finish_reason),
                    logprobs: None,
                }
            })
            .collect();

        // AI21 doesn't provide token counts in the same format
        let usage = Usage {
            prompt_tokens: 0, // Would need to estimate
            completion_tokens: 0,
            total_tokens: 0,
            prompt_tokens_details: None,
            completion_tokens_details: None,
        };

        Ok(ChatCompletionResponse {
            id: format!("chatcmpl-ai21-{}", uuid::Uuid::new_v4()),
            object: "chat.completion".to_string(),
            created: chrono::Utc::now().timestamp() as u64,
            model: model.to_string(),
            choices: choices
                .into_iter()
                .map(|choice| ChatChoice {
                    index: choice.index,
                    message: choice.message,
                    logprobs: choice.logprobs.map(|_| Logprobs { content: None }),
                    finish_reason: choice.finish_reason,
                })
                .collect(),
            usage: Some(usage),
            system_fingerprint: None,
        })
    }
}

#[async_trait]
impl Provider for AI21Provider {
    fn name(&self) -> &str {
        &self.base.name
    }

    fn provider_type(&self) -> ProviderType {
        ProviderType::Custom("ai21".to_string())
    }

    async fn supports_model(&self, model: &str) -> bool {
        self.base.is_model_supported(model) || model.starts_with("j2-") || model.contains("jamba")
    }

    async fn supports_images(&self) -> bool {
        false
    }

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

    async fn supports_streaming(&self) -> bool {
        false // AI21 doesn't support streaming
    }

    async fn list_models(&self) -> Result<Vec<Model>> {
        let known_models = vec!["j2-ultra", "j2-mid", "j2-light", "jamba-instruct"];

        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: "ai21".to_string(),
            })
            .collect();

        Ok(models)
    }

    async fn health_check(&self) -> Result<()> {
        debug!("Performing AI21 health check");
        Ok(()) // AI21 doesn't have a specific health endpoint
    }

    async fn chat_completion(
        &self,
        request: ChatCompletionRequest,
        _context: RequestContext,
    ) -> Result<ChatCompletionResponse> {
        debug!("AI21 chat completion for model: {}", request.model);

        let prompt = self.convert_messages_to_ai21(&request.messages);

        let mut body = json!({
            "prompt": prompt
        });

        // Add optional parameters
        if let Some(max_tokens) = request.max_tokens {
            body["maxTokens"] = json!(max_tokens);
        }
        if let Some(temperature) = request.temperature {
            body["temperature"] = json!(temperature);
        }
        if let Some(top_p) = request.top_p {
            body["topP"] = json!(top_p);
        }

        let url = format!(
            "{}/studio/v1/{}/complete",
            self.base.base_url, request.model
        );

        let response = self
            .base
            .client
            .post(&url)
            .header("Authorization", format!("Bearer {}", self.base.api_key))
            .header("Content-Type", "application/json")
            .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),
                404 => ProviderError::ModelNotFound(error_text),
                400 => ProviderError::InvalidRequest(error_text),
                _ => ProviderError::Unknown(format!("HTTP {}: {}", status, error_text)),
            }
            .into());
        }

        let ai21_response: serde_json::Value = self.base.parse_json_response(response).await?;
        self.convert_ai21_response_to_openai(ai21_response, &request.model)
    }

    async fn completion(
        &self,
        request: CompletionRequest,
        _context: RequestContext,
    ) -> Result<CompletionResponse> {
        debug!("AI21 completion for model: {}", request.model);

        let mut body = json!({
            "prompt": request.prompt
        });

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

        let url = format!(
            "{}/studio/v1/{}/complete",
            self.base.base_url, request.model
        );

        let response = self
            .base
            .client
            .post(&url)
            .header("Authorization", format!("Bearer {}", self.base.api_key))
            .header("Content-Type", "application/json")
            .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),
                404 => ProviderError::ModelNotFound(error_text),
                400 => ProviderError::InvalidRequest(error_text),
                _ => ProviderError::Unknown(format!("HTTP {}: {}", status, error_text)),
            }
            .into());
        }

        let ai21_response: serde_json::Value = self.base.parse_json_response(response).await?;

        let empty_vec = vec![];
        let completions = ai21_response
            .get("completions")
            .and_then(|c| c.as_array())
            .unwrap_or(&empty_vec);

        let choices: Vec<CompletionChoice> = completions
            .iter()
            .enumerate()
            .map(|(index, completion)| {
                let text = completion
                    .get("data")
                    .and_then(|d| d.get("text"))
                    .and_then(|t| t.as_str())
                    .unwrap_or("")
                    .to_string();

                let finish_reason = completion
                    .get("finishReason")
                    .and_then(|fr| fr.get("reason"))
                    .and_then(|r| r.as_str())
                    .map(|r| match r {
                        "stop" => "stop",
                        "length" => "length",
                        _ => "stop",
                    })
                    .unwrap_or("stop")
                    .to_string();

                CompletionChoice {
                    text,
                    index: index as u32,
                    logprobs: None,
                    finish_reason: Some(finish_reason),
                }
            })
            .collect();

        Ok(CompletionResponse {
            id: format!("cmpl-ai21-{}", uuid::Uuid::new_v4()),
            object: "text_completion".to_string(),
            created: chrono::Utc::now().timestamp() as u64,
            model: request.model,
            choices,
            usage: Some(Usage::default()),
        })
    }

    async fn embedding(
        &self,
        _request: EmbeddingRequest,
        _context: RequestContext,
    ) -> Result<EmbeddingResponse> {
        Err(ProviderError::InvalidRequest("Embeddings not supported by AI21".to_string()).into())
    }

    async fn image_generation(
        &self,
        _request: ImageGenerationRequest,
        _context: RequestContext,
    ) -> Result<ImageGenerationResponse> {
        Err(
            ProviderError::InvalidRequest("Image generation not supported by AI21".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 {
            Ok(ModelPricing {
                model: model.to_string(),
                input_cost_per_1k: 0.01,
                output_cost_per_1k: 0.01,
                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)
    }
}