zclaw_openfang/desktop/src-tauri/src/intelligence/extraction_adapter.rs

//! Extraction Adapter - Bridges zclaw_growth::LlmDriverForExtraction with the Kernel's LlmDriver
//!
//! Implements the `LlmDriverForExtraction` trait by delegating to the Kernel's
//! `zclaw_runtime::driver::LlmDriver`, which already handles provider-specific
//! API calls (OpenAI, Anthropic, Gemini, etc.).
//!
//! This enables the Growth system's MemoryExtractor to call the LLM for memory
//! extraction from conversations.

use std::sync::Arc;
use zclaw_growth::extractor::{LlmDriverForExtraction, prompts};
use zclaw_growth::types::{ExtractedMemory, MemoryType};
use zclaw_runtime::driver::{CompletionRequest, ContentBlock, LlmDriver};
use zclaw_types::{Message, Result, SessionId};

/// Adapter that wraps the Kernel's `LlmDriver` to implement `LlmDriverForExtraction`.
///
/// The adapter translates extraction requests into completion requests that the
/// Kernel's LLM driver can process, then parses the structured JSON response
/// back into `ExtractedMemory` objects.
pub struct TauriExtractionDriver {
    driver: Arc<dyn LlmDriver>,
    model: String,
}

impl TauriExtractionDriver {
    /// Create a new extraction driver wrapping the given LLM driver.
    ///
    /// The `model` parameter specifies which model to use for extraction calls.
    pub fn new(driver: Arc<dyn LlmDriver>, model: String) -> Self {
        Self { driver, model }
    }

    /// Build a completion request from the extraction prompt and conversation messages.
    fn build_request(
        &self,
        messages: &[Message],
        extraction_type: MemoryType,
    ) -> CompletionRequest {
        let extraction_prompt = prompts::get_extraction_prompt(extraction_type);

        // Format conversation for the prompt
        // Message is an enum with variants: User{content}, Assistant{content, thinking},
        // System{content}, ToolUse{...}, ToolResult{...}
        let conversation_text = messages
            .iter()
            .filter_map(|msg| {
                match msg {
                    Message::User { content } => {
                        Some(format!("[User]: {}", content))
                    }
                    Message::Assistant { content, .. } => {
                        Some(format!("[Assistant]: {}", content))
                    }
                    Message::System { content } => {
                        Some(format!("[System]: {}", content))
                    }
                    // Skip tool use/result messages -- not relevant for memory extraction
                    Message::ToolUse { .. } | Message::ToolResult { .. } => None,
                }
            })
            .collect::<Vec<_>>()
            .join("\n\n");

        let full_prompt = format!("{}{}", extraction_prompt, conversation_text);

        CompletionRequest {
            model: self.model.clone(),
            system: Some(
                "You are a memory extraction assistant. Analyze conversations and extract \
                structured memories in valid JSON format. Always respond with valid JSON only, \
                no additional text or markdown formatting."
                    .to_string(),
            ),
            messages: vec![Message::user(full_prompt)],
            tools: Vec::new(),
            max_tokens: Some(2000),
            temperature: Some(0.3),
            stop: Vec::new(),
            stream: false,
            thinking_enabled: false,
            reasoning_effort: None,
            plan_mode: false,
        }
    }

    /// Parse the LLM response text into a list of extracted memories.
    fn parse_response(
        &self,
        response_text: &str,
        extraction_type: MemoryType,
    ) -> Vec<ExtractedMemory> {
        // Strip markdown code fences if present
        let cleaned = response_text
            .trim()
            .trim_start_matches("```json")
            .trim_start_matches("```")
            .trim_end_matches("```")
            .trim();

        // Extract the JSON array from the response
        let json_str = match (cleaned.find('['), cleaned.rfind(']')) {
            (Some(start), Some(end)) => &cleaned[start..=end],
            _ => {
                tracing::warn!(
                    "[TauriExtractionDriver] No JSON array found in LLM response"
                );
                return Vec::new();
            }
        };

        let raw_items: Vec<serde_json::Value> = match serde_json::from_str(json_str) {
            Ok(items) => items,
            Err(e) => {
                tracing::warn!(
                    "[TauriExtractionDriver] Failed to parse extraction JSON: {}",
                    e
                );
                return Vec::new();
            }
        };

        raw_items
            .into_iter()
            .filter_map(|item| self.parse_memory_item(&item, extraction_type))
            .collect()
    }

    /// Parse a single memory item from JSON.
    fn parse_memory_item(
        &self,
        value: &serde_json::Value,
        fallback_type: MemoryType,
    ) -> Option<ExtractedMemory> {
        let content = value.get("content")?.as_str()?.to_string();
        let category = value
            .get("category")
            .and_then(|v| v.as_str())
            .unwrap_or("unknown")
            .to_string();

        let confidence = value
            .get("confidence")
            .and_then(|v| v.as_f64())
            .unwrap_or(0.7) as f32;

        let keywords = value
            .get("keywords")
            .and_then(|v| v.as_array())
            .map(|arr| {
                arr.iter()
                    .filter_map(|v| v.as_str().map(String::from))
                    .collect()
            })
            .unwrap_or_default();

        Some(
            ExtractedMemory::new(fallback_type, category, content, SessionId::new())
                .with_confidence(confidence)
                .with_keywords(keywords),
        )
    }
}

#[async_trait::async_trait]
impl LlmDriverForExtraction for TauriExtractionDriver {
    async fn extract_memories(
        &self,
        messages: &[Message],
        extraction_type: MemoryType,
    ) -> Result<Vec<ExtractedMemory>> {
        let type_name = format!("{}", extraction_type);
        tracing::debug!(
            "[TauriExtractionDriver] Extracting {} memories from {} messages",
            type_name,
            messages.len()
        );

        // Skip extraction if there are too few messages
        if messages.len() < 2 {
            tracing::debug!(
                "[TauriExtractionDriver] Too few messages ({}) for extraction, skipping",
                messages.len()
            );
            return Ok(Vec::new());
        }

        let request = self.build_request(messages, extraction_type);

        let response = self.driver.complete(request).await.map_err(|e| {
            tracing::error!(
                "[TauriExtractionDriver] LLM completion failed for {}: {}",
                type_name,
                e
            );
            e
        })?;

        // Extract text content from response
        let response_text: String = response
            .content
            .into_iter()
            .filter_map(|block| match block {
                ContentBlock::Text { text } => Some(text),
                _ => None,
            })
            .collect::<Vec<_>>()
            .join("");

        if response_text.is_empty() {
            tracing::warn!(
                "[TauriExtractionDriver] Empty response from LLM for {} extraction",
                type_name
            );
            return Ok(Vec::new());
        }

        let memories = self.parse_response(&response_text, extraction_type);

        tracing::info!(
            "[TauriExtractionDriver] Extracted {} {} memories",
            memories.len(),
            type_name
        );

        Ok(memories)
    }

    async fn extract_with_prompt(
        &self,
        messages: &[Message],
        system_prompt: &str,
        user_prompt: &str,
    ) -> Result<String> {
        if messages.len() < 2 {
            return Err(zclaw_types::ZclawError::InvalidInput(
                "Too few messages for combined extraction".to_string(),
            ));
        }

        tracing::debug!(
            "[TauriExtractionDriver] Combined extraction from {} messages",
            messages.len()
        );

        let request = CompletionRequest {
            model: self.model.clone(),
            system: Some(system_prompt.to_string()),
            messages: vec![Message::user(user_prompt.to_string())],
            tools: Vec::new(),
            max_tokens: Some(3000),
            temperature: Some(0.3),
            stop: Vec::new(),
            stream: false,
            thinking_enabled: false,
            reasoning_effort: None,
            plan_mode: false,
        };

        let response = self.driver.complete(request).await.map_err(|e| {
            tracing::error!(
                "[TauriExtractionDriver] Combined extraction LLM call failed: {}",
                e
            );
            e
        })?;

        let response_text: String = response
            .content
            .into_iter()
            .filter_map(|block| match block {
                ContentBlock::Text { text } => Some(text),
                _ => None,
            })
            .collect::<Vec<_>>()
            .join("");

        if response_text.is_empty() {
            return Err(zclaw_types::ZclawError::LlmError(
                "Empty response from LLM for combined extraction".to_string(),
            ));
        }

        tracing::info!(
            "[TauriExtractionDriver] Combined extraction response: {} chars",
            response_text.len()
        );

        Ok(response_text)
    }
}

/// Global extraction driver instance (legacy path, kept for compatibility).
///
/// **Architecture note:** The Kernel struct now holds its own `extraction_driver` field
/// (set via `kernel.set_extraction_driver()`), which is the primary path used by
/// the middleware chain. This OnceCell global is a legacy artifact — its accessors
/// are dead code. The `configure_extraction_driver()` function is still called during
/// kernel_init for backward compatibility but the primary consumption path is
/// through the Kernel struct.
static EXTRACTION_DRIVER: tokio::sync::OnceCell<Arc<TauriExtractionDriver>> =
    tokio::sync::OnceCell::const_new();

/// Configure the global extraction driver (legacy path).
///
/// Called during kernel initialization. The primary path is via
/// `kernel.set_extraction_driver()` which stores the driver in the Kernel struct
/// for use by the middleware chain.
pub fn configure_extraction_driver(driver: Arc<dyn LlmDriver>, model: String) {
    let adapter = TauriExtractionDriver::new(driver, model);
    let _ = EXTRACTION_DRIVER.set(Arc::new(adapter));
    tracing::info!("[ExtractionAdapter] Extraction driver configured (legacy OnceCell path)");
}

/// Check if the extraction driver is available (legacy OnceCell path).
#[allow(dead_code)] // @reserved: diagnostic check for legacy OnceCell extraction driver
pub fn is_extraction_driver_configured() -> bool {
    EXTRACTION_DRIVER.get().is_some()
}

/// Get the global extraction driver (legacy OnceCell path).
///
/// Prefer accessing via `kernel.extraction_driver()` when the Kernel is available.
#[allow(dead_code)] // @reserved: legacy accessor, prefer kernel.extraction_driver()
pub fn get_extraction_driver() -> Option<Arc<TauriExtractionDriver>> {
    EXTRACTION_DRIVER.get().cloned()
}

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

    #[test]
    fn test_extraction_driver_not_configured_by_default() {
        assert!(!is_extraction_driver_configured());
    }

    #[test]
    fn test_parse_valid_json_response() {
        let response = r#"```json
        [
            {
                "category": "communication-style",
                "content": "User prefers concise replies",
                "confidence": 0.9,
                "keywords": ["concise", "style"]
            },
            {
                "category": "language",
                "content": "User prefers Chinese responses",
                "confidence": 0.85,
                "keywords": ["Chinese", "language"]
            }
        ]
        ```"#;

        // Verify the parsing logic works by manually simulating it
        let cleaned = response
            .trim()
            .trim_start_matches("```json")
            .trim_start_matches("```")
            .trim_end_matches("```")
            .trim();

        let json_str = &cleaned[cleaned.find('[').unwrap()..=cleaned.rfind(']').unwrap()];
        let items: Vec<serde_json::Value> = serde_json::from_str(json_str).unwrap();
        assert_eq!(items.len(), 2);
        assert_eq!(
            items[0].get("category").unwrap().as_str().unwrap(),
            "communication-style"
        );
    }

    #[test]
    fn test_parse_no_json_array() {
        let response = "No memories could be extracted from this conversation.";
        let has_array =
            response.find('[').is_some() && response.rfind(']').is_some();
        assert!(!has_array);
    }
}