---
stepsCompleted: [1, 2, 3, 4, 5]
workflow_completed: true
inputDocuments:
  - _bmad-output/planning-artifacts/prd-phase1-mvp-ai.md
  - _bmad-output/planning-artifacts/ux-design-specification.md
  - _bmad-output/planning-artifacts/architecture.md
  - _bmad-output/analysis/brainstorming-session-2026-01-09.md
workflow_type: 'create-epics-and-stories'
project_name: 'Keep (Memento Phase 1 MVP AI)'
user_name: 'Ramez'
date: '2026-01-10'
focus_area: 'Phase 1 MVP AI - AI-Powered Note Taking Features'
communication_language: 'French'
document_output_language: 'English'
status: 'completed'
---

# Keep (Memento) - Epic Breakdown - Phase 1 MVP AI

## Overview

This document provides the complete epic and story breakdown for **Keep Phase 1 MVP AI**, decomposing the requirements from the Phase 1 PRD, UX Design Specification, and Architecture into implementable stories.

**Project Context:** Brownfield extension of existing Keep Notes application with AI-powered features. Zero breaking changes to existing functionality.

**Implementation Timeline:** 12 weeks (4 phases)
**Target:** Production-ready MVP with 6 core AI features

---

## Requirements Inventory

### Functional Requirements - Phase 1 MVP

**Core AI Features:**
- **FR6:** Real-time content analysis for concept identification
- **FR7:** AI-powered tag suggestions based on content analysis
- **FR8:** User control over AI suggestions (accept/modify/reject)
- **FR11:** Exact keyword search (title and content)
- **FR12:** Semantic search by meaning/intention (natural language)
- **FR13:** Hybrid search combining exact + semantic results

**Foundation Features (Already Implemented):**
- **FR1:** CRUD operations for notes (text and checklist)
- **FR2:** Pin notes to top of list
- **FR3:** Archive notes
- **FR4:** Attach images to notes
- **FR5:** Drag-and-drop reordering (Muuri)
- **FR9:** Manual tag management
- **FR10:** Filter and sort by tags
- **FR16:** Optimistic UI for immediate feedback

**Configuration & Administration:**
- **FR17:** AI provider configuration (OpenAI, Ollama)
- **FR18:** Multi-provider support via Vercel AI SDK
- **FR19:** Theme customization (dark mode)

**Deferred to Phase 2/3:**
- **FR14:** Offline PWA mode
- **FR15:** Background sync

### Non-Functional Requirements

**Performance:**
- **NFR1:** Auto-tagging < 1.5s after typing ends
- **NFR2:** Semantic search < 300ms for 1000 notes
- **NFR3:** Title suggestions < 2s after detection

**Security & Privacy:**
- **NFR4:** API key isolation (server-side only)
- **NFR5:** Local-first privacy (Ollama = 100% local)

**Reliability:**
- **NFR8:** Vector integrity (automatic background updates)

**Portability:**
- **NFR9:** Minimal footprint (Zero DevOps)
- **NFR10:** Node.js LTS support

---

## Phase 1 MVP AI Epic Mapping

### Epic 1: Intelligent Title Suggestions ⭐
**Focus:** AI-powered title generation for untitled notes
**FRs covered:** FR6, FR8
**Architecture Decision:** Decision 1 (Database Schema), Decision 3 (Language Detection)
**Priority:** HIGH (Core user experience feature)

### Epic 2: Hybrid Semantic Search 🔍
**Focus:** Keyword + vector search with RRF fusion
**FRs covered:** FR11, FR12, FR13
**Architecture Decision:** Decision 1 (Database Schema - reuses Note.embedding)
**Priority:** HIGH (Core discovery feature)

### Epic 3: Paragraph-Level Reformulation ✍️
**Focus:** AI-powered text improvement (Clarify, Shorten, Improve Style)
**FRs covered:** FR6, FR8
**Architecture Decision:** Decision 1 (Database Schema - no schema change)
**Priority:** MEDIUM (User productivity feature)

### Epic 4: Memory Echo (Proactive Connections) 🧠
**Focus:** Daily proactive note connections via cosine similarity
**FRs covered:** FR6
**Architecture Decision:** Decision 2 (Memory Echo Architecture)
**Priority:** HIGH (Differentiating feature)

### Epic 5: AI Settings Panel ⚙️
**Focus:** Granular ON/OFF controls per feature + provider selection
**FRs covered:** FR17, FR18
**Architecture Decision:** Decision 4 (AI Settings Architecture)
**Priority:** HIGH (User control requirement)

### Epic 6: Language Detection Service 🌐
**Focus:** Automatic language detection (TinyLD hybrid approach)
**FRs covered:** FR6 (Cross-cutting concern)
**Architecture Decision:** Decision 3 (Language Detection Strategy)
**Priority:** HIGH (Enables multilingual prompts)

---

## Epic 1: Intelligent Title Suggestions

### Overview
Generate 3 AI-powered title suggestions when a note reaches 50+ words without a title. User can accept, modify, or reject suggestions.

**User Stories:** 3
**Estimated Complexity:** Medium
**Dependencies:** Language Detection Service, AI Provider Factory

### Story 1.1: Real-time Word Count Detection
**As a user, I want the system to detect when my note reaches 50+ words without a title, so that I can receive title suggestions automatically.**

**Acceptance Criteria:**
- **Given** an open note editor
- **When** I type content and the word count reaches 50+
- **And** the note title field is empty
- **Then** the system triggers background title generation
- **And** a non-intrusive toast notification appears: "💡 Title suggestions available"

**Technical Requirements:**
- Word count triggered on `debounce` (300ms after typing stops)
- Detection logic: `content.split(/\s+/).length >= 50`
- Must not interfere with typing experience (non-blocking)
- Toast notification uses Sonner (radix-ui compatible)

**Implementation Files:**
- Component: `keep-notes/components/ai/ai-suggestion.tsx` (NEW)
- Hook: `useWordCountDetection` (NEW utility)
- UI: Toast notification with "View" / "Dismiss" actions

---

### Story 1.2: AI Title Generation
**As a system, I want to generate 3 relevant title suggestions using AI, so that users can quickly organize their notes.**

**Acceptance Criteria:**
- **Given** a note with 50+ words of content
- **When** title generation is triggered
- **Then** the AI generates 3 distinct title suggestions
- **And** each title is concise (3-8 words)
- **And** titles reflect the main concept of the content
- **And** generation completes within < 2 seconds

**Technical Requirements:**
- Service: `TitleSuggestionService` in `lib/ai/services/title-suggestion.service.ts`
- Provider: Uses `getAIProvider()` factory (OpenAI or Ollama)
- System Prompt: English (stability)
- User Data: Local language (FR, EN, ES, DE, FA, etc.)
- Language Detection: Called before generation for multilingual prompts
- Storage: Suggestions stored in memory (not persisted until user accepts)

**Prompt Engineering:**
```
System: You are a title generator. Generate 3 concise titles (3-8 words each) that capture the main concept.
User Language: {detected_language}
Content: {note_content}

Output format: JSON array of strings
```

**Error Handling:**
- If AI fails: Retry once with different provider (if available)
- If retry fails: Show toast error "Failed to generate suggestions. Please try again."
- Timeout: 5 seconds maximum

---

### Story 1.3: User Interaction & Feedback
**As a user, I want to accept, modify, or reject title suggestions, so that I maintain full control over my note organization.**

**Acceptance Criteria:**
- **Given** 3 AI-generated title suggestions
- **When** I click the toast notification
- **Then** a modal displays the 3 suggestions
- **And** I can click any suggestion to apply it as the note title
- **And** I can click "Dismiss" to ignore all suggestions
- **And** the modal closes automatically after selection or dismissal

**Technical Requirements:**
- Component: `AiSuggestionModal` (extends `components/ai/ai-suggestion.tsx`)
- Server Action: `updateNote(noteId, { title })`
- Feedback: Store user choice in `AiFeedback` table
  - `feedbackType`: 'thumbs_up' if accepted without modification
  - `feature`: 'title_suggestion'
  - `originalContent`: All 3 suggestions (JSON array)
  - `correctedContent`: User's final choice (or modified title)

**UI/UX Requirements (from UX Design Spec):**
- Modal design: Clean, centered, with card-style suggestions
- Each suggestion: Clickable card with hover effect
- "Dismiss" button: Secondary action at bottom
- Auto-close after selection (no confirmation dialog)
- If user modifies title: Record as 'correction' feedback

**Implementation Files:**
- Modal: `components/ai/ai-suggestion.tsx` (NEW)
- Server Action: `app/actions/ai-suggestions.ts` (NEW)
- API Route: `/api/ai/feedback` (NEW) - stores feedback

**Database Updates:**
```typescript
// When user accepts a title
await prisma.note.update({
  where: { id: noteId },
  data: {
    title: selectedTitle,
    autoGenerated: true,
    aiProvider: currentProvider,
    aiConfidence: 85, // Placeholder - Phase 3 will calculate
    lastAiAnalysis: new Date()
  }
})

// Store feedback for Phase 3 trust scoring
await prisma.aiFeedback.create({
  data: {
    noteId,
    userId: session.user.id,
    feedbackType: 'thumbs_up',
    feature: 'title_suggestion',
    originalContent: JSON.stringify(allThreeSuggestions),
    correctedContent: selectedTitle,
    metadata: JSON.stringify({
      provider: currentProvider,
      model: modelName,
      timestamp: new Date()
    })
  }
})
```

---

## Epic 2: Hybrid Semantic Search

### Overview
Combine exact keyword matching with vector similarity search using Reciprocal Rank Fusion (RRF) for comprehensive results.

**User Stories:** 3
**Estimated Complexity:** High
**Dependencies:** Existing embeddings system, Language Detection (optional)

### Story 2.1: Query Embedding Generation
**As a system, I want to generate vector embeddings for user search queries, so that I can find notes by meaning.**

**Acceptance Criteria:**
- **Given** a user search query
- **When** the search is executed
- **Then** the system generates a vector embedding for the query
- **And** the embedding is stored in memory (not persisted)
- **And** generation completes within < 200ms

**Technical Requirements:**
- Service: `SemanticSearchService` in `lib/ai/services/semantic-search.service.ts`
- Provider: Uses `getAIProvider()` factory
- Embedding Model: `text-embedding-3-small` (OpenAI) or Ollama equivalent
- Language Detection: Optional (can detect query language for better results)
- Caching: Query embeddings cached in React Cache (5-minute TTL)

**Implementation:**
```typescript
// lib/ai/services/semantic-search.service.ts
async generateQueryEmbedding(query: string): Promise<number[]> {
  const provider = getAIProvider()
  const embedding = await provider.generateEmbedding(query)
  return embedding
}
```

---

### Story 2.2: Vector Similarity Calculation
**As a system, I want to calculate cosine similarity between query and all user notes, so that I can rank results by meaning.**

**Acceptance Criteria:**
- **Given** a query embedding and all user note embeddings
- **When** similarity calculation runs
- **Then** the system calculates cosine similarity for each note
- **And** returns notes ranked by similarity score (descending)
- **And** calculation completes within < 300ms for 1000 notes

**Technical Requirements:**
- Algorithm: Cosine similarity
- Formula: `similarity = dotProduct(queryEmbedding, noteEmbedding) / (magnitude(query) * magnitude(note))`
- Threshold: Notes with similarity < 0.3 are filtered out
- Performance: In-memory calculation (no separate vector DB for Phase 1)

**Implementation:**
```typescript
// lib/ai/services/semantic-search.service.ts
async searchBySimilarity(
  queryEmbedding: number[],
  userId: string
): Promise<Array<{ note: Note, score: number }>> {
  // Fetch all user notes with embeddings
  const notes = await prisma.note.findMany({
    where: { userId },
    select: { id: true, title: true, content: true, embedding: true }
  })

  // Calculate cosine similarity
  const results = notes
    .map(note => ({
      note,
      score: cosineSimilarity(queryEmbedding, JSON.parse(note.embedding))
    }))
    .filter(r => r.score > 0.3) // Threshold filter
    .sort((a, b) => b.score - a.score)

  return results
}
```

---

### Story 2.3: Hybrid Search with RRF Fusion
**As a user, I want to see combined results from keyword search and semantic search, so that I get the most comprehensive results.**

**Acceptance Criteria:**
- **Given** a search query
- **When** I execute the search
- **Then** the system performs BOTH keyword search AND semantic search
- **And** results are fused using Reciprocal Rank Fusion (RRF)
- **And** each result displays a badge: "Exact Match" or "Related"
- **And** total time < 300ms for 1000 notes

**Technical Requirements:**
- Service: `SemanticSearchService` (extends from Story 2.1, 2.2)
- Fusion Algorithm: Reciprocal Rank Fusion (RRF)
  - `RRF(score) = 1 / (k + rank)` where k = 60 (standard value)
  - Combined score = `RRF(keyword_rank) + RRF(semantic_rank)`
- Keyword Search: Existing Prisma query (title/content LIKE `%query%`)
- Semantic Search: Cosine similarity from Story 2.2
- Result Limit: Top 20 notes

**RRF Implementation:**
```typescript
// lib/ai/services/semantic-search.service.ts
async hybridSearch(
  query: string,
  userId: string
): Promise<Array<{ note: Note, keywordScore: number, semanticScore: number, combinedScore: number }>> {
  // Parallel execution
  const [keywordResults, semanticResults] = await Promise.all([
    this.keywordSearch(query, userId),  // Existing implementation
    this.searchBySimilarity(query, userId)  // Story 2.2
  ])

  // Calculate RRF scores
  const k = 60
  const scoredNotes = new Map<string, any>()

  // Add keyword RRF scores
  keywordResults.forEach((note, index) => {
    const rrf = 1 / (k + index + 1)
    scoredNotes.set(note.id, {
      note,
      keywordScore: rrf,
      semanticScore: 0,
      combinedScore: rrf
    })
  })

  // Add semantic RRF scores and combine
  semanticResults.forEach(({ note, score }, index) => {
    const rrf = 1 / (k + index + 1)
    if (scoredNotes.has(note.id)) {
      const existing = scoredNotes.get(note.id)
      existing.semanticScore = rrf
      existing.combinedScore += rrf
    } else {
      scoredNotes.set(note.id, {
        note,
        keywordScore: 0,
        semanticScore: rrf,
        combinedScore: rrf
      })
    }
  })

  // Convert to array and sort by combined score
  return Array.from(scoredNotes.values())
    .sort((a, b) => b.combinedScore - a.combinedScore)
    .slice(0, 20) // Top 20 results
}
```

**UI Requirements (from UX Design Spec):**
- Component: `components/ai/semantic-search-results.tsx` (NEW)
- Badge display:
  - "Exact Match" badge: Blue background, shown if `keywordScore > 0`
  - "Related" badge: Gray background, shown if `semanticScore > 0` AND `keywordScore === 0`
  - Both badges can appear if note matches both
- Result card: Displays title, content snippet (100 chars), badges
- Loading state: Skeleton cards while searching (< 300ms)

**API Route:**
- Endpoint: `POST /api/ai/search`
- Request schema:
  ```typescript
  { query: string, userId: string }
  ```
- Response:
  ```typescript
  {
    success: true,
    data: {
      results: Array<{
        note: Note,
        badges: Array<"Exact Match" | "Related">
      }>,
      totalResults: number,
      searchTime: number // milliseconds
    }
  }
  ```

---

## Epic 3: Paragraph-Level Reformulation

### Overview
AI-powered text improvement with 3 options: Clarify, Shorten, Improve Style. Triggered via context menu on text selection.

**User Stories:** 2
**Estimated Complexity:** Medium
**Dependencies:** AI Provider Factory

### Story 3.1: Context Menu Integration
**As a user, I want to select text and see "Reformulate" options in a context menu, so that I can improve my writing with AI assistance.**

**Acceptance Criteria:**
- **Given** a note editor with text content
- **When** I select one or more paragraphs (50-500 words)
- **And** I right-click or long-press
- **Then** a context menu appears with "Reformulate" submenu
- **And** the submenu shows 3 options: "Clarify", "Shorten", "Improve Style"
- **When** I click any option
- **Then** the selected text is sent to AI for reformulation
- **And** a loading indicator appears on the selected text

**Technical Requirements:**
- Component: `components/ai/paragraph-refactor.tsx` (NEW)
- Context Menu: Extends existing note editor context menu (Radix Dropdown Menu)
- Text Selection: `window.getSelection()` API
- Word Count Validation: 50-500 words (show error if out of range)
- Loading State: Skeleton or spinner overlay on selected text

**UI Implementation:**
```typescript
// components/ai/paragraph-refactor.tsx
'use client'

import { useCallback } from 'react'
import { startTransition } from 'react'

export function ParagraphRefactor({ noteId, content }: { noteId: string, content: string }) {
  const handleTextSelection = useCallback(() => {
    const selection = window.getSelection()
    const selectedText = selection?.toString()
    const wordCount = selectedText?.split(/\s+/).length || 0

    if (wordCount < 50 || wordCount > 500) {
      showError('Please select 50-500 words to reformulate')
      return
    }

    // Show context menu at selection position
    showContextMenu(selection.getRangeAt(0))
  }, [])

  const handleRefactor = async (option: 'clarify' | 'shorten' | 'improve') => {
    const selectedText = window.getSelection()?.toString()

    startTransition(async () => {
      showLoadingState()

      const result = await refactorParagraph(noteId, selectedText, option)

      hideLoadingState()
      showRefactorDialog(result.refactoredText)
    })
  }

  return (
    // Context menu integration
    <DropdownMenu>
      <DropdownMenuTrigger>Reformulate</DropdownMenuTrigger>
      <DropdownMenuContent>
        <DropdownMenuItem onClick={() => handleRefactor('clarify')}>
          Clarify
        </DropdownMenuItem>
        <DropdownMenuItem onClick={() => handleRefactor('shorten')}>
          Shorten
        </DropdownMenuItem>
        <DropdownMenuItem onClick={() => handleRefactor('improve')}>
          Improve Style
        </DropdownMenuItem>
      </DropdownMenuContent>
    </DropdownMenu>
  )
}
```

---

### Story 3.2: AI Reformulation & Application
**As a user, I want to see AI-reformulated text and choose to apply or discard it, so that I can improve my writing while maintaining control.**

**Acceptance Criteria:**
- **Given** selected text sent for reformulation
- **When** AI completes processing (< 2 seconds)
- **Then** a modal displays showing:
  - Original text (left side)
  - Reformulated text (right side) with diff highlighting
  - "Apply" and "Discard" buttons
- **When** I click "Apply"
- **Then** the reformulated text replaces the original in the note
- **And** the change is saved automatically
- **When** I click "Discard"
- **Then** the modal closes and no changes are made

**Technical Requirements:**
- Service: `ParagraphRefactorService` in `lib/ai/services/paragraph-refactor.service.ts`
- Provider: Uses `getAIProvider()` factory
- System Prompt: English (stability)
- User Data: Local language (respects language detection)
- Diff Display: Use `react-diff-viewer` or similar library

**Prompt Engineering:**
```
System: You are a text reformulator. Reformulate the text according to the user's chosen option.
User Language: {detected_language}
Option: {clarify|shorten|improve}

Clarify: Make the text clearer and easier to understand
Shorten: Reduce word count by 30-50% while keeping key information
Improve Style: Enhance readability, flow, and professional tone

Original Text:
{selected_text}

Output: Reformulated text only (no explanations)
```

**UI Implementation:**
```typescript
// Modal component (extends paragraph-refactor.tsx)
export function RefactorModal({
  originalText,
  refactoredText,
  onApply,
  onDiscard
}) {
  return (
    <Dialog open={true}>
      <DialogContent className="max-w-4xl">
        <DialogHeader>
          <DialogTitle>Compare & Apply</DialogTitle>
        </DialogHeader>

        <div className="grid grid-cols-2 gap-4">
          <div>
            <h4 className="font-medium mb-2">Original</h4>
            <div className="p-4 bg-gray-100 rounded">
              {originalText}
            </div>
          </div>

          <div>
            <h4 className="font-medium mb-2">Refactored</h4>
            <div className="p-4 bg-blue-50 rounded">
              {refactoredText}
            </div>
          </div>
        </div>

        <DialogFooter>
          <Button variant="ghost" onClick={onDiscard}>
            Discard
          </Button>
          <Button onClick={onApply}>
            Apply Changes
          </Button>
        </DialogFooter>
      </DialogContent>
    </Dialog>
  )
}
```

**Server Action:**
```typescript
// app/actions/ai-suggestions.ts
'use server'

import { auth } from '@/auth'
import { ParagraphRefactorService } from '@/lib/ai/services/paragraph-refactor.service'
import { updateNote } from './notes'

export async function refactorParagraph(
  noteId: string,
  selectedText: string,
  option: 'clarify' | 'shorten' | 'improve'
) {
  const session = await auth()
  if (!session?.user?.id) throw new Error('Unauthorized')

  const service = new ParagraphRefactorService()
  const refactoredText = await service.refactor(selectedText, option)

  return {
    success: true,
    originalText: selectedText,
    refactoredText
  }
}

export async function applyRefactoring(
  noteId: string,
  originalText: string,
  refactoredText: string
) {
  const session = await auth()
  if (!session?.user?.id) throw new Error('Unauthorized')

  // Get current note content
  const note = await prisma.note.findUnique({ where: { id: noteId } })
  if (!note?.userId || note.userId !== session.user.id) {
    throw new Error('Note not found')
  }

  // Replace original text with refactored text
  const newContent = note.content.replace(originalText, refactoredText)

  await updateNote(noteId, { content: newContent })

  return { success: true }
}
```

**Feedback Collection:**
```typescript
// Track which reformulation option users prefer
await prisma.aiFeedback.create({
  data: {
    noteId,
    userId: session.user.id,
    feedbackType: 'correction', // User chose to apply
    feature: 'paragraph_refactor',
    originalContent: originalText,
    correctedContent: refactoredText,
    metadata: JSON.stringify({
      option, // 'clarify' | 'shorten' | 'improve'
      provider: currentProvider,
      timestamp: new Date()
    })
  }
})
```

---

## Epic 4: Memory Echo (Proactive Connections)

### Overview
Background process that identifies connections between notes using cosine similarity. Displays 1 insight per day (max similarity > 0.75).

**User Stories:** 2
**Estimated Complexity:** High
**Dependencies:** Existing embeddings system, Decision 2 (Server Action + Queue pattern)

### Story 4.1: Background Insight Generation
**As a system, I want to analyze all user note embeddings daily to find connections, so that I can proactively suggest related notes.**

**Acceptance Criteria:**
- **Given** a user with 10+ notes (each with embeddings)
- **When** the user logs in
- **And** no insight has been generated today
- **Then** the system triggers background analysis
- **And** calculates cosine similarity between all note pairs
- **And** finds the top pair with similarity > 0.75
- **And** stores the insight in `MemoryEchoInsight` table
- **And** UI freeze is < 100ms (only DB check, background processing)

**Technical Requirements:**
- Server Action: `app/actions/ai-memory-echo.ts` (NEW)
- Service: `MemoryEchoService` in `lib/ai/services/memory-echo.service.ts` (NEW)
- Trigger: User login check (in layout or dashboard)
- Constraint: Max 1 insight per user per day (enforced via DB unique constraint)
- Performance: < 100ms UI freeze (async processing)

**Implementation:**
```typescript
// app/actions/ai-memory-echo.ts
'use server'

import { auth } from '@/auth'
import { prisma } from '@/lib/prisma'
import { MemoryEchoService } from '@/lib/ai/services/memory-echo.service'

export async function generateMemoryEcho() {
  const session = await auth()
  if (!session?.user?.id) {
    return { success: false, error: 'Unauthorized' }
  }

  // Check if already generated today
  const today = new Date()
  today.setHours(0, 0, 0, 0)

  const existing = await prisma.memoryEchoInsight.findFirst({
    where: {
      userId: session.user.id,
      insightDate: { gte: today }
    }
  })

  if (existing) {
    return { success: true, insight: existing, alreadyGenerated: true }
  }

  // Generate new insight (non-blocking background task)
  generateInBackground(session.user.id)

  // Return immediately (UI doesn't wait)
  return { success: true, insight: null, alreadyGenerated: false }
}

async function generateInBackground(userId: string) {
  const service = new MemoryEchoService()

  try {
    const insight = await service.findTopConnection(userId)

    if (insight) {
      await prisma.memoryEchoInsight.create({
        data: {
          userId,
          note1Id: insight.note1Id,
          note2Id: insight.note2Id,
          similarityScore: insight.score
        }
      })
    }
  } catch (error) {
    console.error('Memory Echo background generation error:', error)
  }
}
```

**Service Implementation:**
```typescript
// lib/ai/services/memory-echo.service.ts
export class MemoryEchoService {
  async findTopConnection(
    userId: string
  ): Promise<{ note1Id: string, note2Id: string, score: number } | null> {
    // Fetch all user notes with embeddings
    const notes = await prisma.note.findMany({
      where: { userId },
      select: { id: true, embedding: true, title: true, content: true }
    })

    if (notes.length < 2) return null

    // Calculate pairwise cosine similarities
    const insights = []
    const threshold = 0.75

    for (let i = 0; i < notes.length; i++) {
      for (let j = i + 1; j < notes.length; j++) {
        const embedding1 = JSON.parse(notes[i].embedding)
        const embedding2 = JSON.parse(notes[j].embedding)
        const similarity = cosineSimilarity(embedding1, embedding2)

        if (similarity > threshold) {
          insights.push({
            note1Id: notes[i].id,
            note2Id: notes[j].id,
            score: similarity
          })
        }
      }
    }

    // Return top insight (highest similarity)
    if (insights.length === 0) return null

    insights.sort((a, b) => b.score - a.score)
    return insights[0]
  }
}

// Cosine similarity utility
function cosineSimilarity(vecA: number[], vecB: number[]): number {
  const dotProduct = vecA.reduce((sum, a, i) => sum + a * vecB[i], 0)
  const magnitudeA = Math.sqrt(vecA.reduce((sum, a) => sum + a * a, 0))
  const magnitudeB = Math.sqrt(vecB.reduce((sum, b) => sum + b * b, 0))
  return dotProduct / (magnitudeA * magnitudeB)
}
```

---

### Story 4.2: Insight Display & Feedback
**As a user, I want to see daily note connections and provide feedback, so that I can discover relationships in my knowledge base.**

**Acceptance Criteria:**
- **Given** a stored Memory Echo insight
- **When** I log in (or navigate to dashboard)
- **Then** a toast notification appears: "💡 Memory Echo: Note X relates to Note Y (85% match)"
- **When** I click the toast
- **Then** a modal displays both notes side-by-side
- **And** I can click each note to view it in editor
- **And** I can provide feedback via 👍 / 👎 buttons
- **When** I click feedback
- **Then** the feedback is stored in `MemoryEchoInsight.feedback` field

**Technical Requirements:**
- Component: `components/ai/memory-echo-notification.tsx` (NEW)
- Trigger: Check on page load (dashboard layout)
- UI: Toast notification with Sonner
- Modal: Side-by-side note comparison
- Feedback: Updates `MemoryEchoInsight.feedback` field

**UI Implementation:**
```typescript
// components/ai/memory-echo-notification.tsx
'use client'

import { useEffect, useState } from 'react'
import { useRouter } from 'next/navigation'
import { toast } from 'sonner'
import { Bell, X, ThumbsUp, ThumbsDown } from 'lucide-react'
import { generateMemoryEcho } from '@/app/actions/ai-memory-echo'

export function MemoryEchoNotification() {
  const router = useRouter()
  const [insight, setInsight] = useState<any>(null)
  const [viewed, setViewed] = useState(false)

  useEffect(() => {
    checkForInsight()
  }, [])

  const checkForInsight = async () => {
    const result = await generateMemoryEcho()

    if (result.success && result.insight && !result.alreadyGenerated) {
      // Show toast notification
      toast('💡 Memory Echo', {
        description: `Note "${insight.note1.title}" relates to "${insight.note2.title}" (${Math.round(insight.similarityScore * 100)}% match)`,
        action: {
          label: 'View',
          onClick: () => showInsightModal(result.insight)
        }
      })
    }

    if (result.success && result.insight) {
      setInsight(result.insight)
    }
  }

  const showInsightModal = (insightData: any) => {
    // Open modal with both notes side-by-side
    setViewed(true)
    markAsViewed(insightData.id)
  }

  const handleFeedback = async (feedback: 'thumbs_up' | 'thumbs_down') => {
    await updateMemoryEchoFeedback(insight.id, feedback)

    toast(feedback === 'thumbs_up' ? 'Thanks for your feedback!' : 'We\'ll improve next time')

    // Close modal or hide toast
  }

  if (!insight) return null

  return (
    // Modal implementation with feedback buttons
    <Dialog open={viewed} onOpenChange={setViewed}>
      <DialogContent>
        <DialogHeader>
          <DialogTitle>Memory Echo Discovery</DialogTitle>
        </DialogHeader>

        <div className="grid grid-cols-2 gap-4">
          {/* Note 1 */}
          <NoteCard note={insight.note1} onClick={() => router.push(`/notes/${insight.note1.id}`)} />

          {/* Note 2 */}
          <NoteCard note={insight.note2} onClick={() => router.push(`/notes/${insight.note2.id}`)} />
        </div>

        <div className="text-center text-sm text-gray-600">
          Similarity: {Math.round(insight.similarityScore * 100)}%
        </div>

        <div className="flex justify-center gap-4">
          <Button
            variant={insight.feedback === 'thumbs_up' ? 'default' : 'outline'}
            size="icon"
            onClick={() => handleFeedback('thumbs_up')}
          >
            <ThumbsUp className="h-4 w-4" />
          </Button>
          <Button
            variant={insight.feedback === 'thumbs_down' ? 'default' : 'outline'}
            size="icon"
            onClick={() => handleFeedback('thumbs_down')}
          >
            <ThumbsDown className="h-4 w-4" />
          </Button>
        </div>
      </DialogContent>
    </Dialog>
  )
}
```

**Server Action for Feedback:**
```typescript
// app/actions/ai-memory-echo.ts
export async function updateMemoryEchoFeedback(
  insightId: string,
  feedback: 'thumbs_up' | 'thumbs_down'
) {
  const session = await auth()
  if (!session?.user?.id) throw new Error('Unauthorized')

  await prisma.memoryEchoInsight.update({
    where: { id: insightId },
    data: { feedback }
  })

  return { success: true }
}
```

**Database Schema (from Architecture Decision 2):**
```prisma
model MemoryEchoInsight {
  id              String    @id @default(cuid())
  userId          String?
  note1Id         String
  note2Id         String
  similarityScore Float
  insightDate     DateTime  @default(now())
  viewed          Boolean   @default(false)
  feedback        String?

  note1           Note      @relation("EchoNote1", fields: [note1Id], references: [id])
  note2           Note      @relation("EchoNote2", fields: [note2Id], references: [id])
  user            User?     @relation(fields: [userId], references: [id])

  @@unique([userId, insightDate])
  @@index([userId, insightDate])
}
```

---

## Epic 5: AI Settings Panel

### Overview
Dedicated settings page at `/settings/ai` with granular ON/OFF controls for each AI feature and provider selection.

**User Stories:** 2
**Estimated Complexity:** Medium
**Dependencies:** Decision 4 (UserAISettings table), AI Provider Factory

### Story 5.1: Granular Feature Toggles
**As a user, I want to enable/disable individual AI features, so that I can control which AI assistance I receive.**

**Acceptance Criteria:**
- **Given** the AI Settings page at `/settings/ai`
- **When** I navigate to the page
- **Then** I see toggles for each AI feature:
  - Title Suggestions (default: ON)
  - Semantic Search (default: ON)
  - Paragraph Reformulation (default: ON)
  - Memory Echo (default: ON)
- **When** I toggle any feature OFF
- **Then** the setting is saved to `UserAISettings` table
- **And** the feature is immediately disabled in the UI
- **When** I toggle any feature ON
- **Then** the feature is re-enabled immediately

**Technical Requirements:**
- Page: `app/(main)/settings/ai/page.tsx` (NEW)
- Component: `components/ai/ai-settings-panel.tsx` (NEW)
- Server Action: `app/actions/ai-settings.ts` (NEW)
- Database: `UserAISettings` table (from Decision 4)

**UI Implementation:**
```typescript
// app/(main)/settings/ai/page.tsx
import { AISettingsPanel } from '@/components/ai/ai-settings-panel'
import { getAISettings } from '@/lib/ai/settings'

export default async function AISettingsPage() {
  const settings = await getAISettings()

  return (
    <div className="container mx-auto py-8">
      <h1 className="text-3xl font-bold mb-6">AI Settings</h1>

      <AISettingsPanel initialSettings={settings} />
    </div>
  )
}

// components/ai/ai-settings-panel.tsx
'use client'

import { useState } from 'react'
import { Switch } from '@/components/ui/switch'
import { Label } from '@/components/ui/label'
import { Card } from '@/components/ui/card'
import { updateAISettings } from '@/app/actions/ai-settings'

export function AISettingsPanel({ initialSettings }: { initialSettings: any }) {
  const [settings, setSettings] = useState(initialSettings)

  const handleToggle = async (feature: string, value: boolean) => {
    // Optimistic update
    setSettings(prev => ({ ...prev, [feature]: value }))

    // Server update
    await updateAISettings({ [feature]: value })
  }

  return (
    <div className="space-y-6">
      <FeatureToggle
        name="titleSuggestions"
        label="Title Suggestions"
        description="Suggest titles for untitled notes after 50+ words"
        checked={settings.titleSuggestions}
        onChange={(checked) => handleToggle('titleSuggestions', checked)}
      />

      <FeatureToggle
        name="semanticSearch"
        label="Semantic Search"
        description="Find notes by meaning, not just keywords"
        checked={settings.semanticSearch}
        onChange={(checked) => handleToggle('semanticSearch', checked)}
      />

      <FeatureToggle
        name="paragraphRefactor"
        label="Paragraph Reformulation"
        description="AI-powered text improvement options"
        checked={settings.paragraphRefactor}
        onChange={(checked) => handleToggle('paragraphRefactor', checked)}
      />

      <FeatureToggle
        name="memoryEcho"
        label="Memory Echo"
        description="Daily proactive connections between your notes"
        checked={settings.memoryEcho}
        onChange={(checked) => handleToggle('memoryEcho', checked)}
      />

      {settings.memoryEcho && (
        <FrequencySlider
          value={settings.memoryEchoFrequency}
          onChange={(value) => handleToggle('memoryEchoFrequency', value)}
          options={['daily', 'weekly', 'custom']}
        />
      )}
    </div>
  )
}

function FeatureToggle({
  name,
  label,
  description,
  checked,
  onChange
}: {
  name: string
  label: string
  description: string
  checked: boolean
  onChange: (checked: boolean) => void
}) {
  return (
    <Card className="p-4">
      <div className="flex items-center justify-between">
        <div className="space-y-1">
          <Label htmlFor={name}>{label}</Label>
          <p className="text-sm text-gray-500">{description}</p>
        </div>
        <Switch
          id={name}
          checked={checked}
          onCheckedChange={onChange}
        />
      </div>
    </Card>
  )
}
```

**Server Action:**
```typescript
// app/actions/ai-settings.ts
'use server'

import { auth } from '@/auth'
import { prisma } from '@/lib/prisma'

export async function updateAISettings(settings: Partial<UserAISettings>) {
  const session = await auth()
  if (!session?.user?.id) throw new Error('Unauthorized')

  // Upsert settings (create if not exists)
  await prisma.userAISettings.upsert({
    where: { userId: session.user.id },
    create: {
      userId: session.user.id,
      ...settings
    },
    update: settings
  })

  revalidatePath('/settings/ai')
  return { success: true }
}

export async function getAISettings() {
  const session = await auth()
  if (!session?.user?.id) {
    // Return defaults for non-logged-in users
    return {
      titleSuggestions: true,
      semanticSearch: true,
      paragraphRefactor: true,
      memoryEcho: true,
      memoryEchoFrequency: 'daily',
      aiProvider: 'auto'
    }
  }

  const settings = await prisma.userAISettings.findUnique({
    where: { userId: session.user.id }
  })

  return settings || {
    titleSuggestions: true,
    semanticSearch: true,
    paragraphRefactor: true,
    memoryEcho: true,
    memoryEchoFrequency: 'daily',
    aiProvider: 'auto'
  }
}
```

---

### Story 5.2: AI Provider Selection
**As a user, I want to choose my AI provider (Auto, OpenAI, or Ollama), so that I can control cost and privacy.**

**Acceptance Criteria:**
- **Given** the AI Settings page
- **When** I scroll to the "AI Provider" section
- **Then** I see 3 provider options:
  - **Auto (Recommended)** - Ollama when available, OpenAI fallback
  - **Ollama (Local)** - 100% private, runs locally
  - **OpenAI (Cloud)** - Most accurate, requires API key
- **When** I select a provider
- **Then** the selection is saved to `UserAISettings.aiProvider`
- **And** the AI provider factory uses my preference

**Technical Requirements:**
- Component: Extends `AISettingsPanel` with provider selector
- Integration: `getAIProvider()` factory respects user selection
- Validation: API key required for OpenAI (stored in SystemConfig)

**UI Implementation:**
```typescript
// components/ai/ai-settings-panel.tsx (extend existing component)

function ProviderSelector({
  value,
  onChange
}: {
  value: 'auto' | 'openai' | 'ollama'
  onChange: (value: 'auto' | 'openai' | 'ollama') => void
}) {
  const providers = [
    {
      value: 'auto',
      label: 'Auto (Recommended)',
      description: 'Ollama when available, OpenAI fallback'
    },
    {
      value: 'ollama',
      label: 'Ollama (Local)',
      description: '100% private, runs locally on your machine'
    },
    {
      value: 'openai',
      label: 'OpenAI (Cloud)',
      description: 'Most accurate, requires API key'
    }
  ]

  return (
    <Card className="p-4">
      <Label className="text-base font-medium">AI Provider</Label>

      <RadioGroup value={value} onValueChange={onChange}>
        {providers.map(provider => (
          <div key={provider.value} className="flex items-start space-x-2 py-2">
            <RadioGroupItem value={provider.value} id={provider.value} />
            <div className="grid gap-1.5">
              <Label htmlFor={provider.value}>{provider.label}</Label>
              <p className="text-sm text-gray-500">{provider.description}</p>
            </div>
          </div>
        ))}
      </RadioGroup>

      {value === 'openai' && (
        <APIKeyInput />
      )}
    </Card>
  )
}
```

**Provider Factory Integration:**
```typescript
// lib/ai/factory.ts (existing, extend to respect user settings)

import { getAIProvider } from './factory'
import { getAISettings } from './settings'

export async function getUserAIProvider(): Promise<AIProvider> {
  const userSettings = await getAISettings()
  const systemConfig = await getSystemConfig()

  let provider = userSettings.aiProvider // 'auto' | 'openai' | 'ollama'

  // Handle 'auto' mode
  if (provider === 'auto') {
    // Check if Ollama is available
    try {
      const ollamaStatus = await checkOllamaHealth()
      provider = ollamaStatus ? 'ollama' : 'openai'
    } catch {
      provider = 'openai' // Fallback to OpenAI
    }
  }

  return getAIProvider(provider)
}
```

**Database Schema (from Decision 4):**
```prisma
model UserAISettings {
  userId                String    @id

  // Feature Flags (granular ON/OFF)
  titleSuggestions      Boolean   @default(true)
  semanticSearch        Boolean   @default(true)
  paragraphRefactor     Boolean   @default(true)
  memoryEcho            Boolean   @default(true)

  // Configuration
  memoryEchoFrequency   String    @default("daily")  // 'daily' | 'weekly' | 'custom'
  aiProvider            String    @default("auto")    // 'auto' | 'openai' | 'ollama'

  // Relation
  user                  User      @relation(fields: [userId], references: [id])

  // Indexes for analytics
  @@index([memoryEcho])
  @@index([aiProvider])
  @@index([memoryEchoFrequency])
}
```

---

## Epic 6: Language Detection Service

### Overview
Automatic language detection using TinyLD (62 languages including Persian). Hybrid approach: TinyLD for < 50 words, AI for ≥ 50 words.

**User Stories:** 2
**Estimated Complexity:** Medium
**Dependencies:** Decision 3 (Language Detection Strategy), TinyLD library

### Story 6.1: TinyLD Integration for Short Notes
**As a system, I want to detect note language efficiently for notes < 50 words using TinyLD, so that I can enable multilingual AI processing.**

**Acceptance Criteria:**
- **Given** a note with < 50 words
- **When** the note is saved or analyzed
- **Then** the system detects language using TinyLD
- **And** detection completes in < 10ms
- **And** the detected language is stored in `Note.language` field
- **And** confidence score is stored in `Note.languageConfidence` field

**Technical Requirements:**
- Library: `tinyld` (npm install tinyld)
- Service: `LanguageDetectionService` in `lib/ai/services/language-detection.service.ts`
- Supported Languages: 62 (including Persian/fa verified)
- Output Format: ISO 639-1 codes (fr, en, es, de, fa, etc.)

**Implementation:**
```typescript
// lib/ai/services/language-detection.service.ts
import { tinyld } from 'tinyld'

export class LanguageDetectionService {
  private readonly MIN_WORDS_FOR_AI = 50
  private readonly MIN_CONFIDENCE = 0.7

  async detectLanguage(content: string): Promise<{
    language: string      // 'fr' | 'en' | 'es' | 'de' | 'fa' | 'unknown'
    confidence: number   // 0.0-1.0
    method: 'tinyld' | 'ai' | 'manual'
  }> {
    const wordCount = content.split(/\s+/).length

    // Short notes: TinyLD (fast, TypeScript native)
    if (wordCount < this.MIN_WORDS_FOR_AI) {
      const result = tinyld(content)
      return {
        language: this.mapToISO(result.language),
        confidence: result.confidence || 0.8,
        method: 'tinyld'
      }
    }

    // Long notes: AI for better accuracy
    const response = await generateText({
      model: openai('gpt-4o-mini'), // or ollama/llama3.2
      prompt: `Detect the language of this text. Respond ONLY with ISO 639-1 code (fr, en, es, de, fa):\n\n${content.substring(0, 500)}`
    })

    return {
      language: response.text.toLowerCase().trim(),
      confidence: 0.9,
      method: 'ai'
    }
  }

  private mapToISO(code: string): string {
    const mapping = {
      'fra': 'fr',
      'eng': 'en',
      'spa': 'es',
      'deu': 'de',
      'fas': 'fa',
      'pes': 'fa',  // Persian (Farsi)
      'por': 'pt',
      'ita': 'it',
      'rus': 'ru',
      'zho': 'zh'
    }
    return mapping[code] || code.substring(0, 2)
  }
}
```

**Trigger Points:**
1. Note creation (on save)
2. Note update (on save)
3. Before AI processing (title generation, reformulation, etc.)

**Database Update:**
```typescript
// app/actions/notes.ts (extend existing createNote/updateNote)

export async function createNote(data: { title: string, content: string }) {
  const session = await auth()
  if (!session?.user?.id) throw new Error('Unauthorized')

  // Detect language
  const languageService = new LanguageDetectionService()
  const { language, languageConfidence } = await languageService.detectLanguage(data.content)

  const note = await prisma.note.create({
    data: {
      ...data,
      userId: session.user.id,
      language,
      languageConfidence
    }
  })

  return note
}
```

---

### Story 6.2: AI Fallback for Long Notes
**As a system, I want to use AI language detection for notes ≥ 50 words, so that I can achieve higher accuracy for longer content.**

**Acceptance Criteria:**
- **Given** a note with ≥ 50 words
- **When** the note is saved or analyzed
- **Then** the system detects language using AI (OpenAI or Ollama)
- **And** detection completes in < 500ms
- **And** the detected language is stored in `Note.language` field
- **And** confidence score is 0.9 (AI is more accurate)

**Technical Requirements:**
- Provider: Uses `getAIProvider()` factory
- Model: `gpt-4o-mini` (OpenAI) or `llama3.2` (Ollama)
- Prompt: Minimal (only language detection)
- Output: ISO 639-1 code only

**AI Prompt (from Story 6.1):**
```
Detect the language of this text. Respond ONLY with ISO 639-1 code (fr, en, es, de, fa):

{content (first 500 chars)}
```

**Performance Target:**
- TinyLD detection: ~8ms for < 50 words ✅
- AI detection: ~200-500ms for ≥ 50 words ✅
- Overall impact: Negligible for UX

---

## Implementation Phases

### Phase 1: Foundation (Week 1-2)
**Goal:** Database schema and base infrastructure

**Stories:**
- Epic 1-6: All Prisma migrations (3 new tables, extend Note model)
- Epic 6: Language Detection Service (TinyLD integration)
- Epic 5: AI Settings page + UserAISettings table

**Deliverables:**
- ✅ Prisma migrations created and applied
- ✅ `LanguageDetectionService` implemented
- ✅ `/settings/ai` page functional
- ✅ Base AI service layer structure created

---

### Phase 2: Infrastructure (Week 3-4)
**Goal:** Core services and AI provider integration

**Stories:**
- Epic 1: Title Suggestion Service
- Epic 2: Semantic Search Service (part 1 - embeddings)
- Epic 3: Paragraph Refactor Service
- Epic 4: Memory Echo Service (part 1 - background job)

**Deliverables:**
- ✅ All AI services implemented
- ✅ Provider factory extended for new services
- ✅ Server actions created for all features
- ✅ Integration tests passing

---

### Phase 3: AI Features (Week 5-9)
**Goal:** UI components and user-facing features

**Stories:**
- Epic 1: Title Suggestions UI (Stories 1.1, 1.2, 1.3)
- Epic 2: Semantic Search UI (Stories 2.1, 2.2, 2.3)
- Epic 3: Paragraph Reformulation UI (Stories 3.1, 3.2)
- Epic 4: Memory Echo UI (Stories 4.1, 4.2)

**Deliverables:**
- ✅ All AI components implemented
- ✅ Toast notifications working
- ✅ Modals and dialogs functional
- ✅ Feedback collection active

---

### Phase 4: Polish & Testing (Week 10-12)
**Goal:** Quality assurance and performance optimization

**Stories:**
- Epic 1-6: E2E Playwright tests
- Epic 1-6: Performance testing and optimization
- Epic 1-6: Multi-language testing (FR, EN, ES, DE, FA)
- Epic 1-6: Bug fixes and refinement

**Deliverables:**
- ✅ E2E test coverage for all AI features
- ✅ Performance targets met (search < 300ms, titles < 2s, Memory Echo < 100ms UI freeze)
- ✅ Multi-language verification complete
- ✅ Production deployment ready

---

## Dependencies & Critical Path

### Critical Path Implementation
```
Prisma Migrations → Language Detection Service → AI Settings Page
                                        ↓
                                   All AI Services
                                        ↓
                                   UI Components
                                        ↓
                                   Testing & Polish
```

### Parallel Development Opportunities
- **Week 1-2:** Language Detection + AI Settings (independent)
- **Week 3-4:** All AI services (can be developed in parallel)
- **Week 5-9:** UI components (can be developed in parallel per epic)
- **Week 10-12:** Testing (all features tested together)

### Cross-Epic Dependencies
- **All Epics → Epic 6 (Language Detection):** Must detect language before AI processing
- **All Epics → Epic 5 (AI Settings):** Must check feature flags before executing
- **Epic 2 (Semantic Search) → Existing Embeddings:** Reuses `Note.embedding` field
- **Epic 4 (Memory Echo) → Epic 2 (Semantic Search):** Uses cosine similarity from Epic 2

---

## Definition of Done

### Per Story
- [ ] Code implemented following `project-context.md` rules
- [ ] TypeScript strict mode compliance
- [ ] Server actions have `'use server'` directive
- [ ] Components have `'use client'` directive (if interactive)
- [ ] All imports use `@/` alias
- [ ] Error handling with `try/catch` and `console.error()`
- [ ] API responses follow `{success, data, error}` format
- [ ] `auth()` check in all server actions
- [ ] `revalidatePath('/')` after mutations
- [ ] E2E Playwright test written
- [ ] Manual testing completed

### Per Epic
- [ ] All stories completed
- [ ] Integration tests passing
- [ ] Performance targets met
- [ ] User acceptance criteria validated
- [ ] Documentation updated

### Phase 1 MVP AI
- [ ] All 6 epics completed
- [ ] Zero breaking changes to existing features
- [ ] All NFRs met (performance, security, privacy)
- [ ] Multi-language verified (FR, EN, ES, DE, FA)
- [ ] Production deployment ready
- [ ] User feedback collected and analyzed

---

*Generated: 2026-01-10*
*Author: Winston (Architect Agent) with Create Epics & Stories workflow*
*Based on: PRD Phase 1 MVP AI + UX Design Spec + Architecture (2784 lines)*
*Status: READY FOR IMPLEMENTATION*