#!/bin/bash
# AGI API-Enhanced Training
# Uses API to generate high-quality training data for natural language understanding

set -e

# Source environment
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
source "$SCRIPT_DIR/train_init.sh"

# API Configuration
API_ENDPOINT="https://api.anthropic.com/v1/messages"
API_MODEL="claude-3-5-sonnet-20241022"

# Paths
CHAT_BIN="$AGI_HOME/core/build/bin/agi_chat"
API_TRAINING_DATA="$TRAINING_DATA_DIR/api_training_data.jsonl"
ENHANCED_CORPUS="$TRAINING_DATA_DIR/enhanced_corpus.txt"

mkdir -p "$TRAINING_DATA_DIR"

log "───────────────────────────────────────────────"
log "🤖 API-ENHANCED TRAINING PIPELINE"
log "───────────────────────────────────────────────"
log "API Endpoint: $API_ENDPOINT"
log "Model: $API_MODEL"
log "Credits Available: $((MAX_CREDITS - CREDITS_USED))"
log ""

# Function to call Claude API
call_api() {
    local prompt="$1"
    local response_file="$2"

    # Construct API request
    local request_json=$(cat <<EOF
{
  "model": "$API_MODEL",
  "max_tokens": 1024,
  "messages": [
    {
      "role": "user",
      "content": "$prompt"
    }
  ]
}
EOF
)

    # Call API
    local http_code=$(curl -s -w "%{http_code}" -o "$response_file" \
        -X POST "$API_ENDPOINT" \
        -H "Content-Type: application/json" \
        -H "x-api-key: $API_KEY" \
        -H "anthropic-version: 2023-06-01" \
        -d "$request_json")

    if [ "$http_code" = "200" ]; then
        CREDITS_USED=$((CREDITS_USED + 1))
        log "   API call successful (Credits: $CREDITS_USED/$MAX_CREDITS)"
        return 0
    else
        log "   ⚠️  API call failed (HTTP $http_code)"
        return 1
    fi
}

# Training topics for generating Q&A pairs
declare -a TRAINING_TOPICS=(
    "Explain how distributed agent systems work in simple terms"
    "What is Fibonacci growth control in agent-based systems?"
    "Describe MAX consensus algorithm for distributed goal convergence"
    "How does binary protocol improve agent migration efficiency?"
    "Explain natural language understanding in conversational AI"
    "What is the role of memory in AI conversation systems?"
    "Describe knowledge base management in AGI systems"
    "How do hash-based phrase matching algorithms work?"
    "Explain the concept of agent swarms in distributed computing"
    "What are the benefits of P2P architecture in AGI systems?"
)

log "Generating training data using Claude API..."
log "Topics to explore: ${#TRAINING_TOPICS[@]}"
log ""

# Generate enhanced training data
for i in "${!TRAINING_TOPICS[@]}"; do
    TOPIC="${TRAINING_TOPICS[$i]}"
    SESSION=$((i + 1))

    log "Session $SESSION: Generating knowledge for \"${TOPIC:0:50}...\""

    # Check credit limit
    if [ $CREDITS_USED -ge $MAX_CREDITS ]; then
        log "⚠️  Credit limit reached ($CREDITS_USED/$MAX_CREDITS)"
        break
    fi

    # Generate Q&A pair using API
    RESPONSE_FILE=$(mktemp)
    PROMPT="Generate a clear, concise explanation (2-3 sentences) for this AGI system topic: $TOPIC"

    if call_api "$PROMPT" "$RESPONSE_FILE"; then
        # Extract response text from JSON using jq
        ANSWER=$(jq -r '.content[0].text' "$RESPONSE_FILE" 2>/dev/null || echo "API response parse error")

        if [ "$ANSWER" != "API response parse error" ] && [ -n "$ANSWER" ]; then
            # Add to training corpus
            echo "Q: $TOPIC" >> "$ENHANCED_CORPUS"
            echo "A: $ANSWER" >> "$ENHANCED_CORPUS"
            echo "" >> "$ENHANCED_CORPUS"

            # Log to JSONL for future use (escape quotes in answer for valid JSON)
            ESCAPED_ANSWER=$(echo "$ANSWER" | sed 's/"/\\"/g' | tr '\n' ' ')
            echo "{\"topic\":\"$TOPIC\",\"answer\":\"$ESCAPED_ANSWER\",\"timestamp\":\"$(date -u +%Y-%m-%dT%H:%M:%SZ)\"}" >> "$API_TRAINING_DATA"

            log "   Generated: ${ANSWER:0:80}..."
        else
            log "   Failed to parse API response"
        fi
    else
        log "   Skipped due to API error"
    fi

    rm -f "$RESPONSE_FILE"

    # Rate limiting
    sleep 1
done

log ""
log "───────────────────────────────────────────────"
log "✅ API-Enhanced Training Complete"
log "───────────────────────────────────────────────"

# Statistics
if [ -f "$ENHANCED_CORPUS" ]; then
    CORPUS_LINES=$(wc -l < "$ENHANCED_CORPUS")
    CORPUS_WORDS=$(wc -w < "$ENHANCED_CORPUS")
    CORPUS_SIZE=$(du -h "$ENHANCED_CORPUS" | cut -f1)

    log ""
    log "Enhanced Training Data Statistics:"
    log "  Lines: $CORPUS_LINES"
    log "  Words: $CORPUS_WORDS"
    log "  Size: $CORPUS_SIZE"
    log "  File: $ENHANCED_CORPUS"
fi

log ""
log "API Usage:"
log "  Credits Used: $CREDITS_USED"
log "  Credits Remaining: $((MAX_CREDITS - CREDITS_USED))"
log "  Total Budget: $MAX_CREDITS"

# Test the AGI with new knowledge
log ""
log "Testing AGI with enhanced knowledge..."
log ""

TEST_PROMPT="what is MAX consensus?"
log "Test query: $TEST_PROMPT"

RESPONSE=$(echo -e "$TEST_PROMPT\nexit" | "$CHAT_BIN" 2>/dev/null | grep "^AGI:" | head -1 | sed 's/^AGI: //')
log "AGI Response: $RESPONSE"

log ""
log "───────────────────────────────────────────────"
log "🎓 Training pipeline ready for next phase"
log "───────────────────────────────────────────────"
log ""
log "Files created:"
log "  - Enhanced corpus: $ENHANCED_CORPUS"
log "  - API training data: $API_TRAINING_DATA"
log ""
log "To continue training: $0"
log "To chat interactively: $CHAT_BIN"
log ""