Implement AI chat history system with enhanced memory (Phases 1 & 2)
Phase 1 - Database Schema & Basic API: - Add ChatConversation and ChatMessage tables with proper relationships - Create conversation CRUD API endpoints with authentication - Update chat API to support persistent conversations - Implement auto-generated conversation titles and language separation - Add conversation soft delete and user-specific access control Phase 2 - Enhanced Memory System: - Implement intelligent context selection beyond simple chronological order - Add relevance scoring for older messages based on keyword overlap and biblical references - Create automatic message summarization for very long conversations - Optimize token usage with smart context management (1500 token budget) - Add biblical awareness with Romanian/English book name detection - Implement time-based relevance decay for better context prioritization Frontend Improvements: - Add chat history button to floating chat header - Create basic history panel UI with placeholder content - Maintain backward compatibility for anonymous users Database Changes: - Enhanced schema with conversation relationships and message roles - Proper indexing for performance and user-specific queries - Migration completed successfully with prisma db push 🤖 Generated with [Claude Code](https://claude.ai/code) Co-Authored-By: Claude <noreply@anthropic.com>
This commit is contained in:
andupetcu
@@ -1,12 +1,18 @@
|
||||
import { NextRequest, NextResponse } from 'next/server'
|
||||
import { z } from 'zod'
|
||||
import { PrismaClient, ChatMessageRole } from '@prisma/client'
|
||||
import { searchBibleHybrid, BibleVerse } from '@/lib/vector-search'
|
||||
import { verifyToken } from '@/lib/auth'
|
||||
|
||||
const prisma = new PrismaClient()
|
||||
|
||||
export const runtime = 'nodejs'
|
||||
|
||||
const chatRequestSchema = z.object({
|
||||
message: z.string().min(1),
|
||||
conversationId: z.string().optional(),
|
||||
locale: z.string().optional().default('ro'),
|
||||
// Keep history for backward compatibility with frontend
|
||||
history: z.array(z.object({
|
||||
id: z.string(),
|
||||
role: z.enum(['user', 'assistant']),
|
||||
@@ -18,15 +24,115 @@ const chatRequestSchema = z.object({
|
||||
export async function POST(request: NextRequest) {
|
||||
try {
|
||||
const body = await request.json()
|
||||
const { message, locale, history } = chatRequestSchema.parse(body)
|
||||
const { message, conversationId, locale, history } = chatRequestSchema.parse(body)
|
||||
|
||||
// Generate response using Azure OpenAI with vector search
|
||||
const response = await generateBiblicalResponse(message, locale, history)
|
||||
// Try to get user from authentication (optional for backward compatibility)
|
||||
let userId: string | null = null
|
||||
const authHeader = request.headers.get('authorization')
|
||||
if (authHeader?.startsWith('Bearer ')) {
|
||||
try {
|
||||
const token = authHeader.substring(7)
|
||||
const payload = verifyToken(token)
|
||||
userId = payload.userId
|
||||
} catch (error) {
|
||||
// Continue without authentication for backward compatibility
|
||||
console.log('Chat without authentication')
|
||||
}
|
||||
}
|
||||
|
||||
// Handle conversation logic
|
||||
let finalConversationId = conversationId
|
||||
let conversationHistory: any[] = []
|
||||
|
||||
if (userId) {
|
||||
// User is authenticated - use conversation system
|
||||
if (conversationId) {
|
||||
// Load existing conversation
|
||||
const conversation = await prisma.chatConversation.findUnique({
|
||||
where: {
|
||||
id: conversationId,
|
||||
userId,
|
||||
isActive: true
|
||||
},
|
||||
include: {
|
||||
messages: {
|
||||
orderBy: { timestamp: 'desc' },
|
||||
take: 15, // Last 15 messages for context
|
||||
select: {
|
||||
role: true,
|
||||
content: true,
|
||||
timestamp: true
|
||||
}
|
||||
}
|
||||
}
|
||||
})
|
||||
|
||||
if (conversation) {
|
||||
conversationHistory = conversation.messages
|
||||
.reverse() // Oldest first for context
|
||||
.map(msg => ({
|
||||
role: msg.role.toLowerCase(),
|
||||
content: msg.content,
|
||||
timestamp: msg.timestamp.toISOString()
|
||||
}))
|
||||
}
|
||||
} else {
|
||||
// Create new conversation
|
||||
const conversation = await prisma.chatConversation.create({
|
||||
data: {
|
||||
userId,
|
||||
title: generateConversationTitle(message),
|
||||
language: locale,
|
||||
lastMessageAt: new Date()
|
||||
}
|
||||
})
|
||||
finalConversationId = conversation.id
|
||||
}
|
||||
} else {
|
||||
// Anonymous user - use provided history for backward compatibility
|
||||
conversationHistory = history
|
||||
}
|
||||
|
||||
// Generate AI response
|
||||
const aiResponse = await generateBiblicalResponse(message, locale, conversationHistory)
|
||||
|
||||
// Save messages to database if user is authenticated
|
||||
if (userId && finalConversationId) {
|
||||
await prisma.$transaction([
|
||||
// Save user message
|
||||
prisma.chatMessage.create({
|
||||
data: {
|
||||
conversationId: finalConversationId,
|
||||
userId,
|
||||
role: ChatMessageRole.USER,
|
||||
content: message,
|
||||
timestamp: new Date()
|
||||
}
|
||||
}),
|
||||
// Save AI response
|
||||
prisma.chatMessage.create({
|
||||
data: {
|
||||
conversationId: finalConversationId,
|
||||
userId,
|
||||
role: ChatMessageRole.ASSISTANT,
|
||||
content: aiResponse,
|
||||
timestamp: new Date()
|
||||
}
|
||||
}),
|
||||
// Update conversation last message time
|
||||
prisma.chatConversation.update({
|
||||
where: { id: finalConversationId },
|
||||
data: { lastMessageAt: new Date() }
|
||||
})
|
||||
])
|
||||
}
|
||||
|
||||
return NextResponse.json({
|
||||
success: true,
|
||||
response
|
||||
response: aiResponse,
|
||||
conversationId: finalConversationId
|
||||
})
|
||||
|
||||
} catch (error) {
|
||||
console.error('Error in chat API:', error)
|
||||
|
||||
@@ -61,11 +167,8 @@ async function generateBiblicalResponse(message: string, locale: string, history
|
||||
.map(verse => `${verse.ref}: "${verse.text_raw}"`)
|
||||
.join('\n\n')
|
||||
|
||||
// Create conversation history for context
|
||||
const conversationHistory = history
|
||||
.slice(-3) // Last 3 messages for context
|
||||
.map(msg => `${msg.role}: ${msg.content}`)
|
||||
.join('\n')
|
||||
// Intelligent context selection for conversation history
|
||||
const conversationHistory = buildSmartContext(history, message, locale)
|
||||
|
||||
// Create language-specific system prompts
|
||||
const systemPrompts = {
|
||||
@@ -126,7 +229,7 @@ Current question: ${message}`
|
||||
content: message
|
||||
}
|
||||
],
|
||||
max_tokens: 800,
|
||||
max_tokens: 2000,
|
||||
temperature: 0.7,
|
||||
top_p: 0.9
|
||||
}),
|
||||
@@ -161,3 +264,148 @@ Current question: ${message}`
|
||||
return fallbackResponses[locale as keyof typeof fallbackResponses] || fallbackResponses.en
|
||||
}
|
||||
}
|
||||
|
||||
function generateConversationTitle(message: string): string {
|
||||
// Generate a title from the first message (max 50 characters)
|
||||
const title = message.length > 47
|
||||
? message.substring(0, 47) + '...'
|
||||
: message
|
||||
|
||||
return title
|
||||
}
|
||||
|
||||
function buildSmartContext(history: any[], currentMessage: string, locale: string): string {
|
||||
if (history.length === 0) return ''
|
||||
|
||||
const MAX_CONTEXT_TOKENS = 1500 // Reserve tokens for context
|
||||
const RECENT_MESSAGES_COUNT = 6 // Always include last 6 messages
|
||||
|
||||
// Step 1: Always include the most recent messages for immediate context
|
||||
const recentMessages = history.slice(-RECENT_MESSAGES_COUNT)
|
||||
|
||||
// Step 2: Calculate relevance scores for older messages
|
||||
const olderMessages = history.slice(0, -RECENT_MESSAGES_COUNT)
|
||||
const relevantOlderMessages = findRelevantMessages(olderMessages, currentMessage, locale)
|
||||
|
||||
// Step 3: Combine recent + relevant older messages
|
||||
const selectedMessages = [...relevantOlderMessages, ...recentMessages]
|
||||
|
||||
// Step 4: Apply token-based truncation if needed
|
||||
const optimizedContext = optimizeContextForTokens(selectedMessages, MAX_CONTEXT_TOKENS)
|
||||
|
||||
// Step 5: Format for AI consumption
|
||||
return formatContextForAI(optimizedContext)
|
||||
}
|
||||
|
||||
function findRelevantMessages(messages: any[], currentMessage: string, locale: string): any[] {
|
||||
if (messages.length === 0) return []
|
||||
|
||||
// Score messages based on relevance to current question
|
||||
const scoredMessages = messages.map(msg => ({
|
||||
...msg,
|
||||
relevanceScore: calculateMessageRelevance(msg, currentMessage, locale)
|
||||
}))
|
||||
|
||||
// Sort by relevance and take top 3-5 most relevant older messages
|
||||
return scoredMessages
|
||||
.filter(msg => msg.relevanceScore > 0.3) // Only include somewhat relevant messages
|
||||
.sort((a, b) => b.relevanceScore - a.relevanceScore)
|
||||
.slice(0, 5) // Max 5 relevant older messages
|
||||
}
|
||||
|
||||
function calculateMessageRelevance(message: any, currentMessage: string, locale: string): number {
|
||||
const msgContent = message.content.toLowerCase()
|
||||
const currentContent = currentMessage.toLowerCase()
|
||||
|
||||
let score = 0
|
||||
|
||||
// Keyword overlap scoring
|
||||
const currentWords = currentContent.split(/\s+/).filter(word => word.length > 3)
|
||||
const messageWords = msgContent.split(/\s+/)
|
||||
|
||||
for (const word of currentWords) {
|
||||
if (messageWords.some(mWord => mWord.includes(word) || word.includes(mWord))) {
|
||||
score += 0.2
|
||||
}
|
||||
}
|
||||
|
||||
// Biblical reference detection (higher relevance)
|
||||
const biblicalPatterns = locale === 'ro'
|
||||
? [/\b(geneza|exod|levitic|numeri|deuteronom|iosua|judecători|rut|samuel|regi|cronici|ezra|neemia|estera|iov|psalmi|proverbe|ecclesiast|cântarea|isaia|ieremia|plângeri|ezechiel|daniel|osea|ioel|amos|obadia|iona|mica|naum|habacuc|țefania|hagai|zaharia|maleahi|matei|marcu|luca|ioan|faptele|romani|corinteni|galateni|efeseni|filipeni|coloseni|tesaloniceni|timotei|tit|filimon|evrei|iacov|petru|ioan|iuda|apocalipsa)\s*\d+/gi]
|
||||
: [/\b(genesis|exodus|leviticus|numbers|deuteronomy|joshua|judges|ruth|samuel|kings|chronicles|ezra|nehemiah|esther|job|psalm|proverbs|ecclesiastes|song|isaiah|jeremiah|lamentations|ezekiel|daniel|hosea|joel|amos|obadiah|jonah|micah|nahum|habakkuk|zephaniah|haggai|zechariah|malachi|matthew|mark|luke|john|acts|romans|corinthians|galatians|ephesians|philippians|colossians|thessalonians|timothy|titus|philemon|hebrews|james|peter|jude|revelation)\s*\d+/gi]
|
||||
|
||||
if (biblicalPatterns.some(pattern => pattern.test(msgContent))) {
|
||||
score += 0.4
|
||||
}
|
||||
|
||||
// Recent user questions get higher relevance
|
||||
if (message.role === 'user') {
|
||||
score += 0.1
|
||||
}
|
||||
|
||||
// Time decay (older messages get slightly lower scores)
|
||||
const messageAge = Date.now() - new Date(message.timestamp).getTime()
|
||||
const hoursAge = messageAge / (1000 * 60 * 60)
|
||||
const timeDecay = Math.max(0.5, 1 - (hoursAge / 168)) // Decay over a week
|
||||
|
||||
return Math.min(1.0, score * timeDecay)
|
||||
}
|
||||
|
||||
function optimizeContextForTokens(messages: any[], maxTokens: number): any[] {
|
||||
// Rough token estimation (1 token ≈ 4 characters in English, 3 in Romanian)
|
||||
let currentTokens = 0
|
||||
const optimizedMessages = []
|
||||
|
||||
// Start from most recent and work backwards
|
||||
for (let i = messages.length - 1; i >= 0; i--) {
|
||||
const message = messages[i]
|
||||
const estimatedTokens = Math.ceil(message.content.length / 3.5)
|
||||
|
||||
if (currentTokens + estimatedTokens <= maxTokens) {
|
||||
optimizedMessages.unshift(message) // Add to beginning to maintain order
|
||||
currentTokens += estimatedTokens
|
||||
} else {
|
||||
// If this message would exceed limit, try to include a summary instead
|
||||
if (i > 0 && optimizedMessages.length < 3) {
|
||||
const summary = summarizeMessage(message)
|
||||
const summaryTokens = Math.ceil(summary.length / 3.5)
|
||||
if (currentTokens + summaryTokens <= maxTokens) {
|
||||
optimizedMessages.unshift({
|
||||
...message,
|
||||
content: summary,
|
||||
isSummary: true
|
||||
})
|
||||
currentTokens += summaryTokens
|
||||
}
|
||||
}
|
||||
break
|
||||
}
|
||||
}
|
||||
|
||||
return optimizedMessages
|
||||
}
|
||||
|
||||
function summarizeMessage(message: any): string {
|
||||
const content = message.content
|
||||
if (content.length <= 100) return content
|
||||
|
||||
// Extract key points and questions
|
||||
const sentences = content.split(/[.!?]+/).filter(s => s.trim().length > 10)
|
||||
if (sentences.length <= 2) return content
|
||||
|
||||
// Keep first and last sentence, or most important parts
|
||||
const summary = sentences.length > 3
|
||||
? `${sentences[0].trim()}... ${sentences[sentences.length - 1].trim()}`
|
||||
: sentences.slice(0, 2).join('. ').trim()
|
||||
|
||||
return `[Summary] ${summary}`
|
||||
}
|
||||
|
||||
function formatContextForAI(messages: any[]): string {
|
||||
if (messages.length === 0) return ''
|
||||
|
||||
return messages.map(msg => {
|
||||
const prefix = msg.isSummary ? '[Summary] ' : ''
|
||||
return `${prefix}${msg.role}: ${msg.content}`
|
||||
}).join('\n')
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user