digicraft-be/services/geminiService.ts

import { GoogleGenAI } from "@google/genai";
import dotenv from "dotenv";

dotenv.config();

const ai = new GoogleGenAI({ apiKey: process.env.GEMINI_API_KEY as string });

interface StickerPlan {
    prompts: string[];
    characterCore: string;
}

export const geminiService = {
    /**
     * The Brain: Analyzes the user's prompt and generates a list of distinct scenarios.
     * @param originalPrompt The user's raw input
     * @param count Number of variations to generate
     * @returns List of prompts and the core character description
     */
    async generateStickerSetPlan(originalPrompt: string, count: number): Promise<StickerPlan> {
        console.log(`[GeminiService] Planning ${count} sticker variations...`);

        const planningPrompt = `
        You are an expert Sticker Set Planner.
        Analyze this sticker prompt: "${originalPrompt}"
        
        1. Extract the CORE VISUAL IDENTITY (Character description, style, colors, clothes). 
           This must be preserved EXACTLY in every variation.
        
        2. Generate ${count} DISTINCT variation prompts. 
           Each variation must be a different pose, emotion, or action suitable for a sticker pack.
           (e.g., Happy, Sad, Thinking, Coffee, Running, Sleeping, Winking, etc.)
        
        Output JSON ONLY:
        {
          "characterCore": "The extracted core visual description...",
          "variations": [
             "Full prompt for variation 1...",
             "Full prompt for variation 2..."
          ]
        }
        
        Rules for Variations:
        - Combine the "Character Core" with the new pose/action.
        - Ensure the output is a FULL stable diffusion style prompt (tags).
        - Keep the style tags consistent.
        `;

        try {
            const response = await ai.models.generateContent({
                model: "gemini-2.0-flash",
                contents: [{ text: planningPrompt }],
                config: {
                    responseMimeType: "application/json"
                } as any
            });

            const text = response.candidates?.[0]?.content?.parts?.[0]?.text;
            if (!text) throw new Error("No response from Gemini");

            const json = JSON.parse(text);

            let prompts = json.variations || [];

            // SECURITY: Force array to be string[]
            if (!Array.isArray(prompts)) {
                prompts = [originalPrompt];
            }

            // CRITICAL FIX: Ensure we have exactly 'count' items
            // If AI returns fewer, we pad with the existing ones (round-robin)
            if (prompts.length < count) {
                console.warn(`[GeminiService] AI returned only ${prompts.length}/${count} variations. Padding...`);
                while (prompts.length < count) {
                    // pushing a random existing prompt to fill the gap
                    prompts.push(prompts[prompts.length % prompts.length]);
                }
            }

            // If AI determines more are needed, trim? No, extra is fine, but let's slice just in case UI expects N
            // Actually, extra is bonus. But user paid for N? We usually don't charge per variant in this model, but batch cost.

            return {
                prompts: prompts.slice(0, count), // Ensure exact count
                characterCore: json.characterCore || originalPrompt
            };

        } catch (error: any) {
            console.error("Gemini Planning Error:", error);
            // Fallback: Just repeat the prompt if planning fails? Or error out?
            // For now, return the original prompt repeated to avoid crash, but log error.
            return {
                prompts: Array(count).fill(originalPrompt),
                characterCore: originalPrompt
            };
        }
    },

    /**
     * X-Ray: Updates the user on the visual DNA, strategic gaps, and superior prompt.
     */
    async analyzeCompetitorProduct(params: {
        title: string;
        description: string;
        imageBase64: string;
        apiKey?: string;
    }): Promise<any> {
        console.log(`[GeminiService] Running Competitor X-Ray Analysis...`);
        const { title, description, imageBase64, apiKey } = params;

        // Strip header if present to avoid 500 errors
        const cleanBase64 = imageBase64.replace(/^data:image\/\w+;base64,/, "");

        // Use the specialized Vision model for analysis
        const MODEL_NAME = "gemini-2.0-flash";

        const analysisPrompt = `
        You are an elite E-commerce Strategist and Art Director.
        Analyze this competitor product to help us create something SUPERIOR.

        PRODUCT CONTEXT:
        Title: "${title}"
        Description: "${description.substring(0, 500)}..."

        TASK:
        1. VISUAL DNA: Deconstruct the aesthetic formula (Color palette info, composition style, emotional triggers).
        2. SENTIMENT GAP: Identify what is missing or could be improved (e.g., "Lighting is too flat", "Composition is cluttered").
        3. SUPERIOR PROMPT: Write a "Nano Banana" style stable diffusion prompt to generate a version of this product that is 10x BETTER.
           - Must use (weighted:1.2) tags.
           - Must include quality boosters.
           - Must solve the identified gaps.

        OUTPUT JSON ONLY:
        {
          "visualDna": ["Tag 1", "Tag 2", "Hex Colors", "Composition Rule"],
          "sentimentGap": "Brief strategic analysis of weaknesses...",
          "superiorPrompt": "(masterpiece:1.4), ...",
          "gapAnalysis": "Detailed explanation of why the new prompt is better"
        }
        `;

        try {
            const client = apiKey ? new GoogleGenAI({ apiKey }) : ai;

            const response = await client.models.generateContent({
                model: MODEL_NAME,
                contents: [
                    {
                        role: "user",
                        parts: [
                            { text: analysisPrompt },
                            {
                                inlineData: {
                                    mimeType: "image/jpeg",
                                    data: cleanBase64
                                }
                            }
                        ]
                    }
                ],
                config: {
                    responseMimeType: "application/json"
                } as any
            });

            const text = response.candidates?.[0]?.content?.parts?.[0]?.text;
            if (!text) throw new Error("No response from AI");

            return JSON.parse(text);

        } catch (error: any) {
            console.error("Gemini X-Ray Error:", error);
            throw new Error("Failed to analyze product: " + error.message);
        }
    },

    /**
     * Neuro-Scorecard: Analyzes an image for commercial potential using neuro-marketing principles.
     */
    async analyzeImageNeuroScore(params: {
        imageBase64: string;
        apiKey?: string;
    }): Promise<any> {
        console.log(`[GeminiService] Running Neuro-Scorecard Analysis...`);
        const { imageBase64, apiKey } = params;
        // Strip header if present
        const cleanBase64 = imageBase64.replace(/^data:image\/\w+;base64,/, "");

        // Use Vision model for image analysis
        const MODEL_NAME = "gemini-2.0-flash";

        const analysisPrompt = `
        You are a Neuromarketing Expert and Senior Art Director.
        Score this image based on its potential to sell on Etsy.

        ANALYZE THESE DIMENSIONS (Score 0-10):
        1. **Dopamine Hit**: Does it create immediate excitement/craving?
        2. **Serotonin Flow**: Does it evoke trust, calm, or belonging?
        3. **Cognitive Ease**: Is it easy to process instantly? (High score = distinct subject, clear lighting).
        4. **Commercial Fit**: Does it look like a high-end product vs. an amateur photo?

        CRITICAL INSTRUCTION FOR "IMPROVEMENTS":
        - YOU MUST PROVIDE EXACTLY 2 SPECIFIC IMPROVEMENTS PER CATEGORY.
        - EVEN IF THE SCORE IS 10/10, suggest experimental tweaks.
        - Do NOT give generic advice like "increase contrast".
        - BE SPECIFIC to the image content. Mention specific objects, colors, or areas.
        
        ⚠️ FORBIDDEN SUGGESTIONS (NEVER SUGGEST THESE):
        - Mockups (e.g., "show this in a living room", "place on a wall")
        - Context/environment changes (e.g., "add a frame", "show as wall art")
        - Marketing/presentation ideas (e.g., "include in a bundle", "show lifestyle shot")
        
        ✅ ALLOWED SUGGESTIONS (ONLY THESE TYPES):
        - Color adjustments (saturation, hue, warmth, vibrancy)
        - Lighting changes (exposure, shadows, highlights, contrast)
        - Composition tweaks (crop, reframe, balance, focal point)
        - Detail enhancements (sharpness, texture, remove artifacts)
        - Style refinements (artistic filters, mood adjustments, grain)
        
        - Example GOOD: "Increase the saturation of the red vase to make it pop."
        - Example GOOD: "Add subtle vignette to draw eye to center."
        - Example BAD: "Put this in a living room mockup." ❌
        - NEVER RETURN AN EMPTY ARRAY.

        OUTPUT JSON ONLY:
        {
          "scores": {
            "dopamine": 8.5,
            "serotonin": 7.0,
            "cognitiveEase": 9.0,
            "commercialFit": 6.5
          },
          "feedback": [
            "Positive point 1",
            "Positive point 2"
          ],
          "improvements": {
            "dopamine": ["Specific fix 1", "Specific fix 2"],
            "serotonin": ["Specific fix 1", "Specific fix 2"],
            "cognitiveEase": ["Specific fix 1", "Specific fix 2"],
            "commercialFit": ["Specific fix 1", "Specific fix 2"]
          },
          "prediction": "High/Medium/Low Conversion Potential"
        }
        `;

        try {
            const client = apiKey ? new GoogleGenAI({ apiKey }) : ai;

            const response = await client.models.generateContent({
                model: MODEL_NAME,
                contents: [
                    {
                        role: "user",
                        parts: [
                            { text: analysisPrompt },
                            {
                                inlineData: {
                                    mimeType: "image/jpeg",
                                    data: cleanBase64
                                }
                            }
                        ]
                    }
                ],
                config: {
                    responseMimeType: "application/json"
                } as any
            });

            const text = response.candidates?.[0]?.content?.parts?.[0]?.text;
            if (!text) throw new Error("No response from AI");

            const data = JSON.parse(text);

            // NORMALIZE DATA: Check if AI returned old format or missing keys
            if (!data.improvements) {
                console.warn("[GeminiService] AI returned old format, normalizing...");
                data.improvements = {
                    dopamine: [],
                    serotonin: [],
                    cognitiveEase: [],
                    commercialFit: data.criticalImprovements || [] // Fallback to old field
                };
            }

            // Ensure all keys exist
            const defaults = { dopamine: [], serotonin: [], cognitiveEase: [], commercialFit: [] };
            data.improvements = { ...defaults, ...data.improvements };

            return data;

        } catch (error: any) {
            console.error("Gemini Neuro-Score Error:", error);
            throw new Error("Failed to score image: " + error.message);
        }
    },

    /**
     * Web Research: Uses Google Search Grounding to extract metadata from a URL.
     * Bypasses local IP blocking by using Google's servers.
     */
    async performWebResearch(url: string, apiKey?: string): Promise<{ title: string, description: string, image: string }> {
        console.log(`[GeminiService] Performing Google Search Grounding for: ${url}`);

        const researchPrompt = `
        Analyze this product URL: "${url}"
        
        TASK:
        Extract the following metadata from this product page:
        1. Product Title (Exact full title)
        2. Product Description (First 2-3 sentences summary)
        3. Main Product Image URL (Direct link to the highest resolution image, must be a full URL starting with https://)

        IMPORTANT: You MUST return valid JSON. Do not include any text before or after the JSON.

        OUTPUT FORMAT (JSON ONLY):
        {
            "title": "Product Title Here",
            "description": "Product Description Here...",
            "image": "https://full-url-to-image.jpg"
        }
        `;

        try {
            const client = apiKey ? new GoogleGenAI({ apiKey }) : ai;

            // First try with Google Search grounding
            let response;
            try {
                response = await client.models.generateContent({
                    model: "gemini-2.0-flash",
                    contents: [{ text: researchPrompt }],
                    tools: [{
                        googleSearch: {}
                    }],
                    config: {
                        responseMimeType: "application/json"
                    }
                } as any);
            } catch (searchError: any) {
                console.warn("[GeminiService] Google Search grounding failed, trying without:", searchError.message);
                // Fallback: Try without search grounding
                response = await client.models.generateContent({
                    model: "gemini-2.0-flash",
                    contents: [{ text: researchPrompt }],
                    config: {
                        responseMimeType: "application/json"
                    } as any
                });
            }

            const text = response.candidates?.[0]?.content?.parts?.[0]?.text;
            console.log("[GeminiService] Raw research response:", text?.substring(0, 200));

            if (!text) throw new Error("No response from AI Research");

            // Try to parse JSON, handle cases where response might have extra text
            let data;
            try {
                data = JSON.parse(text);
            } catch (parseError) {
                // Try to extract JSON from text
                const jsonMatch = text.match(/\{[\s\S]*\}/);
                if (jsonMatch) {
                    data = JSON.parse(jsonMatch[0]);
                } else {
                    throw new Error("Could not parse JSON from response");
                }
            }

            // Validate required fields
            if (!data.title) {
                console.warn("[GeminiService] Missing title in response, extracting from URL...");
                // Extract title from URL as fallback
                const urlParts = url.split('/');
                const slug = urlParts.find(p => p.length > 10 && !p.includes('.'));
                data.title = slug ? slug.replace(/-/g, ' ') : "Unknown Product";
            }

            if (!data.image || !data.image.startsWith('http')) {
                console.warn("[GeminiService] Invalid or missing image URL");
                data.image = "";
            }

            console.log(`[GeminiService] Research Success: ${data.title}`);
            return data;

        } catch (error: any) {
            console.error("Gemini Research Error:", error);
            throw new Error("Failed to research url: " + error.message);
        }
    }
};