Automated content filling using neural networks for 1C-Bitrix

AI-Powered Content Generation for 1C-Bitrix

GPT integration into Bitrix's content pipeline solves a scale problem: writing unique texts for 50,000 products manually is not feasible, and template-based generation produces predictably weak results. AI generation creates diverse content from structured product data — with the right tone, length, and SEO optimization.

What Can Be Automated with AI

• Product descriptions — unique text based on product attributes
• SEO tags — title, description, keywords
• Section headings — H1, H2 for category pages
• FAQ blocks — frequently asked questions for product cards
• Button text and microcopy — calls to action, hints
• Translations — when content exists in one language

Integration with the OpenAI API

A minimal client for the Chat Completions API:

class OpenAiClient {
    private string $apiKey;
    private string $model = 'gpt-4o-mini';

    public function generate(string $prompt, int $maxTokens = 500): string {
        $response = (new \GuzzleHttp\Client())->post(
            'https://api.openai.com/v1/chat/completions',
            [
                'headers' => ['Authorization' => "Bearer {$this->apiKey}", 'Content-Type' => 'application/json'],
                'json' => [
                    'model' => $this->model,
                    'messages' => [['role' => 'user', 'content' => $prompt]],
                    'max_tokens' => $maxTokens,
                ],
            ]
        );
        return json_decode($response->getBody(), true)['choices'][0]['message']['content'];
    }
}

Cost management: GPT-4o-mini costs ~$0.00015 per 1K input tokens. One product description ≈ 200 prompt tokens + 300 response tokens. 10,000 descriptions ≈ $5. GPT-4o is 10× more expensive but delivers significantly better quality.

Prompt Design

Output quality is determined by the prompt. Structure of an effective product description prompt:

You are a copywriter for an electronics online store.
Write a product description in 2–3 paragraphs (150–200 words) for the following product:
Name: {NAME}
Brand: {BRAND}
Specifications: {SPECS_LIST}

Requirements:
- Style: professional, no hyperbole
- First paragraph — main benefit
- Second paragraph — technical specs in usage context
- Third paragraph — who this product is for
- No phrases like "high quality", "excellent choice"
- Language: English

Prompts are stored in a Highload block AiPrompts linked to the product category — different categories require different styles.

Queue System and Rate Limiting

OpenAI enforces limits: 10,000 RPM and 10,000,000 TPM for GPT-4o-mini. Large catalogs require a queue:

CREATE TABLE ai_generation_queue (
    id SERIAL PRIMARY KEY,
    element_id INT NOT NULL,
    task_type VARCHAR(50),  -- 'description', 'seo_title', 'faq'
    status VARCHAR(20) DEFAULT 'pending',
    result TEXT,
    tokens_used INT,
    error TEXT,
    created_at TIMESTAMP DEFAULT NOW()
);

The worker processes no more than 100 requests per minute, inserting pauses between batches.

Quality Control and Moderation

AI can generate irrelevant or incorrect content. Quality control system:

Automated checks:

• Minimum text length (< 50 characters → error)
• Absence of prohibited words/phrases
• Hallucination check — mentions of attributes not passed in the prompt

Manual review flags: items with a low quality score (determined by a second AI request — prompt "Rate the quality of this description on a scale of 1–10 and give a reason") are flagged for manager review.

Project Timeline

Total: 6–10 working days. Prompt iteration continues for another 1–2 weeks after launch.