package memory

import (
	"context"
	"encoding/json"
	"fmt"
	"log"
	"strings"

	"git.gobha.me/xcaliber/chat-switchboard/database"
	"git.gobha.me/xcaliber/chat-switchboard/knowledge"
	"git.gobha.me/xcaliber/chat-switchboard/models"
	"git.gobha.me/xcaliber/chat-switchboard/providers"
	"git.gobha.me/xcaliber/chat-switchboard/roles"
	"git.gobha.me/xcaliber/chat-switchboard/store"
)

// defaultExtractionPrompt is used when the persona has no custom prompt.
const defaultExtractionPrompt = `You are a memory extraction assistant. Analyze the following conversation and extract memorable facts about the user.

Focus on:
- Personal preferences (language, tools, frameworks, communication style)
- Technical details (deployment stack, database choices, architecture patterns)
- Project context (project names, team size, deadlines, goals)
- Biographical facts (role, company, location, experience level)

Respond ONLY with a JSON array of extracted facts. Each fact must have:
- "key": short label (2-5 words, lowercase)
- "value": the detail/fact (1-2 sentences max)
- "confidence": 0.0-1.0 how certain you are

Example:
[
  {"key": "preferred language", "value": "Go with Gin framework for backend services", "confidence": 0.95},
  {"key": "deployment target", "value": "Kubernetes on AWS with PostgreSQL databases", "confidence": 0.8}
]

If no memorable facts are found, respond with an empty array: []`

// maxConversationChars limits the conversation text sent for extraction.
const maxConversationChars = 30000

// minMessagesForExtraction is the minimum new messages before extraction triggers.
const minMessagesForExtraction = 6

// extractedFact is the JSON structure returned by the utility model.
type extractedFact struct {
	Key        string  `json:"key"`
	Value      string  `json:"value"`
	Confidence float64 `json:"confidence"`
}

// Extractor analyzes conversations and extracts memorable facts.
type Extractor struct {
	stores       store.Stores
	roleResolver *roles.Resolver
	embedder     *knowledge.Embedder
}

// NewExtractor creates a new memory extraction service.
func NewExtractor(stores store.Stores, rr *roles.Resolver, embedder *knowledge.Embedder) *Extractor {
	return &Extractor{stores: stores, roleResolver: rr, embedder: embedder}
}

// Extract analyzes a conversation and saves extracted facts as pending_review memories.
func (e *Extractor) Extract(ctx context.Context, channelID, userID, teamID, personaID string) error {
	if e.stores.Memories == nil || e.stores.Messages == nil {
		return fmt.Errorf("memory or message store not available")
	}

	// Check extraction log for last processed message
	var lastMessageID string
	row := database.DB.QueryRowContext(ctx,
		database.Q(`SELECT last_message_id FROM memory_extraction_log WHERE channel_id = $1 AND user_id = $2`),
		channelID, userID)
	row.Scan(&lastMessageID) // ignore error — may not exist yet

	// Load recent messages for this channel
	messages, err := e.stores.Messages.ListForChannel(ctx, channelID, store.ListOptions{Limit: 200})
	if err != nil {
		return fmt.Errorf("load messages: %w", err)
	}

	// Filter to new messages only (messages come back newest-first)
	var newMessages []models.Message
	for i := len(messages) - 1; i >= 0; i-- {
		if lastMessageID == "" || messages[i].ID > lastMessageID {
			newMessages = append(newMessages, messages[i])
		}
	}

	if len(newMessages) < minMessagesForExtraction {
		return nil // not enough new content
	}

	// Build conversation text
	var sb strings.Builder
	for _, m := range newMessages {
		role := m.Role
		if role == "user" {
			role = "User"
		} else {
			role = "Assistant"
		}
		line := fmt.Sprintf("[%s]: %s\n", role, m.Content)
		if sb.Len()+len(line) > maxConversationChars {
			break
		}
		sb.WriteString(line)
	}

	if sb.Len() < 200 {
		return nil // too little content
	}

	// Get extraction prompt (persona-specific or default)
	prompt := defaultExtractionPrompt
	if personaID != "" && e.stores.Personas != nil {
		persona, err := e.stores.Personas.GetByID(ctx, personaID)
		if err == nil {
			// Respect persona memory_enabled flag (H6 — audit v0.28.0)
			if !persona.MemoryEnabled {
				return nil
			}
			if persona.MemoryExtractionPrompt != nil && *persona.MemoryExtractionPrompt != "" {
				prompt = *persona.MemoryExtractionPrompt
			}
		}
	}

	// Call utility model via role resolver
	var tID *string
	if teamID != "" {
		tID = &teamID
	}

	apiMessages := []providers.Message{
		{Role: "system", Content: prompt},
		{Role: "user", Content: "Analyze this conversation and extract memorable facts:\n\n" + sb.String()},
	}

	result, err := e.roleResolver.Complete(ctx, "utility", userID, tID, apiMessages)
	if err != nil {
		return fmt.Errorf("extraction completion: %w", err)
	}

	// Parse response
	facts, err := parseExtractionResponse(result.Content)
	if err != nil {
		log.Printf("⚠ memory extraction parse failed for channel %s: %v", channelID, err)
		return nil // don't fail on parse errors
	}

	if len(facts) == 0 {
		log.Printf("🧠 memory extraction: channel %s → 0 facts (nothing memorable)", channelID)
	}

	// Determine scope
	scope := models.MemoryScopeUser
	ownerID := userID
	var memUserID *string
	if personaID != "" {
		scope = models.MemoryScopePersonaUser
		ownerID = personaID
		memUserID = &userID
	}

	// Save extracted facts
	saved := 0
	for _, f := range facts {
		if f.Key == "" || f.Value == "" || f.Confidence < 0.3 {
			continue
		}

		mem := &models.Memory{
			ID:              store.NewID(),
			Scope:           scope,
			OwnerID:         ownerID,
			UserID:          memUserID,
			Key:             f.Key,
			Value:           f.Value,
			SourceChannelID: &channelID,
			Confidence:      f.Confidence,
			Status:          models.MemoryStatusPendingReview,
		}

		if err := e.stores.Memories.Upsert(ctx, mem); err != nil {
			log.Printf("⚠ memory save failed: %v", err)
			continue
		}

		// Embed if available
		e.embedMemory(ctx, mem, userID, tID)
		saved++
	}

	// Update extraction log
	latestID := newMessages[len(newMessages)-1].ID
	if database.IsSQLite() {
		_, err = database.DB.ExecContext(ctx, `
			INSERT INTO memory_extraction_log (id, channel_id, user_id, last_message_id, memory_count)
			VALUES (?, ?, ?, ?, ?)
			ON CONFLICT(channel_id, user_id) DO UPDATE SET
				last_message_id = excluded.last_message_id,
				extracted_at = datetime('now'),
				memory_count = memory_extraction_log.memory_count + excluded.memory_count
		`, store.NewID(), channelID, userID, latestID, saved)
	} else {
		_, err = database.DB.ExecContext(ctx, `
			INSERT INTO memory_extraction_log (channel_id, user_id, last_message_id, memory_count)
			VALUES ($1, $2, $3, $4)
			ON CONFLICT(channel_id, user_id) DO UPDATE SET
				last_message_id = EXCLUDED.last_message_id,
				extracted_at = now(),
				memory_count = memory_extraction_log.memory_count + EXCLUDED.memory_count
		`, channelID, userID, latestID, saved)
	}

	if saved > 0 {
		log.Printf("✅ memory extraction: channel %s → %d facts", channelID, saved)
	}
	return err
}

// embedMemory generates and stores an embedding vector for a memory.
func (e *Extractor) embedMemory(ctx context.Context, m *models.Memory, userID string, teamID *string) {
	if e.embedder == nil || !e.embedder.IsConfigured(ctx) {
		return
	}

	text := m.Key + ": " + m.Value
	result, err := e.embedder.EmbedChunks(ctx, userID, teamID, []string{text})
	if err != nil || len(result.Vectors) == 0 {
		return
	}

	vecJSON, _ := json.Marshal(result.Vectors[0])
	if database.IsSQLite() {
		database.DB.ExecContext(ctx,
			`UPDATE memories SET embedding = ? WHERE id = ?`,
			string(vecJSON), m.ID)
	} else {
		database.DB.ExecContext(ctx,
			`UPDATE memories SET embedding = $1::vector WHERE id = $2`,
			string(vecJSON), m.ID)
	}
}

// parseExtractionResponse parses the utility model's JSON response.
func parseExtractionResponse(content string) ([]extractedFact, error) {
	content = strings.TrimSpace(content)

	// Strip markdown code fences
	if strings.HasPrefix(content, "```") {
		lines := strings.Split(content, "\n")
		if len(lines) > 2 {
			lines = lines[1 : len(lines)-1]
			content = strings.Join(lines, "\n")
		}
	}

	var facts []extractedFact
	if err := json.Unmarshal([]byte(content), &facts); err != nil {
		return nil, fmt.Errorf("parse facts JSON: %w", err)
	}
	return facts, nil
}