AI Content Generation: filling fields, properties, and SEO for information blocks using AI

What this solution does

AI content generation module for 1C-Bitrix. The neural network automatically populates product descriptions, section texts, SEO meta tags, and translations directly within infoblocks. The user creates a single rule—a sequence of steps with prompts, field substitution from product cards, and response checks—which the module then applies to tens of thousands of elements in the background. The solution works with keys from Polza.ai, OpenRouter, or a custom OpenAI-compatible gateway. Generated results can be sent to an editor for confirmation and reverted with a single click.

The Problem

In catalogs with several thousand items, content often lags behind product availability. Product cards imported from 1C contain names and article numbers, but descriptions are written irregularly, meta tags remain empty, and category texts are not created at all. These texts are not visible in the admin panel, yet they are responsible for search engine traffic. Hiring a copywriter for large volumes is expensive, while manual writing distracts from store management.

The manual method of copying product cards to a neural network chat and returning the text does not scale: it functions for dozens of items but fails with hundreds. The lack of repeatability means formulations vary with each attempt. There is no control over the output format: the model may respond in English, use Markdown wrapping, or generate text exceeding field limits. Additionally, there is no reporting: it is impossible to track processing status, request costs, or how to revert changes if the result is unsatisfactory.

The module transforms this one-off process into an automated rule. You configure the generation parameters and target fields once, after which the rule applies to any number of elements and sections within the information block.

A rule is a profile comprising a neural network connection, card filters, and a pipeline of sequential steps. Each step performs a single model request and saves the response to a specific field. For instance, the first step generates a full product description, the second creates a short preview based on it, and the third generates meta-title and meta-description from the prepared text. Since each subsequent step utilizes the output of the previous one, there is no need to resend the entire card, optimizing costs.

The module generates product descriptions, populates infoblock properties, and creates SEO texts using artificial intelligence. Prompt templates utilize variables (such as {{Name}}, {{Brand}}, {{Material}}) that are automatically replaced with data from the specific product card, including standard fields, infoblock properties, dictionary values, and section information. The processing logic excludes parts of the prompt corresponding to missing product characteristics, preventing empty fields like 'Brand:' without a value from appearing in the final text.

Processing occurs in the background via the Bitrix queue rather than in the browser. This allows the admin panel to be closed, the server to be restarted, or work to be interrupted; the process will resume from the point of interruption upon the next launch. Before sending the first paid request, the system verifies pipeline configuration settings and the provider account balance. During operation, detailed logging is maintained, recording every prompt sent, every model response, and the cost of each step.

This solution is suitable for online stores with large catalogs where descriptions exist for only a portion of products, content studios managing multiple Bitrix projects, website owners requiring SEO texts for hundreds of sections, and developers tasked with integrating neural networks who wish to avoid writing the integration code from scratch.

How does this differ from a standard AI chat? Through repeatability and control. A rule is configured once to deliver consistent results across the entire catalog; responses are validated before saving, not after; every record is visible in history and can be rolled back.

Capabilities

Neural Network Connection

The module supports Polza.ai, OpenRouter, or a custom endpoint. It does not sell tokens; it operates using your own key. Polza.ai allows payment in rubles without a foreign card or VPN. OpenRouter offers the largest model catalog, though access for users in Russia is currently restricted by the service provider. Any OpenAI-compatible gateway is suitable if data must remain within the internal perimeter. The model catalog with prices is fetched automatically, and a mini-chat is available nearby to verify the key.

Web Search for Specific Steps

Before generating a response, the model can search the internet and write based on found pages rather than relying solely on memory. Settings allow configuring the number of results, a predefined search query, and lists of allowed and blocked domains—for example, restricting to manufacturer websites only and excluding marketplaces. Cited sources are visible in the execution history.

Generation Rule

The module implements a pipeline of sequential generation steps. Each step is configured individually with its own prompt, target field, write mode (overwrite, append, or fill only empty cells), selected model, timeout, and attempt limit. Steps can be temporarily disabled without deleting their configuration.

Prompts are built with dynamic card data substitution. Interface elements allow inserting fields, properties, dictionaries, SEO templates, sections, and results from previous steps via click-to-insert, using search by human-readable field names. Conditional blocks display prompt fragments only when the corresponding field contains data.

Prompt library. Effective formulations are saved to a shared collection with tags and descriptions, enabling reuse across any profile. This prevents loss of work when staff changes.

AI prompt improvement. A dedicated function rewrites the current text for greater precision and structure or generates a prompt from scratch based on a natural language task description. The response is returned in the language of the original input.

Processed objects

The module processes elements and sections of information blocks. It reads standard fields, properties, section custom fields, section data, and SEO templates, then writes the generated result to standard fields, properties, or any available SEO template. Automating text and SEO for categories is often a complex task that is frequently deferred.

The card filter allows setting conditions based on fields, properties, custom fields, and SEO parameters (such as missing meta-description). Filtering can be limited to sections, including those with subsections. The interface immediately displays the number of selected cards, providing an estimate of the workload.

Mass processing is handled via a background queue and agent. The entire current filter can be added to the queue with a single click. The agent processes tasks in batches, remains stable during server restarts, and displays status counts for pending, completed, and failed tasks. Individual cards or groups can be restarted or removed from the queue.

Automatic event capture for information blocks triggers the generation of descriptions and meta tags when a new item appears, without manual intervention. Tasks are created only for cards matching the filter profile. Protection against infinite loops during data rewrites is implemented.

The module allows testing generation on a single card without writing to the database. In debug mode, the final prompt after substitution, the model response, the raw response before processing, validation results, token count, cost, and processing time are displayed. Multiplying the cost of a single run by the number of items in the selection allows you to calculate the campaign budget in advance.

Before the first paid request, pipeline settings are automatically checked: empty prompts for active steps and links to non-existent or disabled stages are detected. The same validation is performed before launching any queue.

Result quality is ensured by stripping technical wrappers from the response. The module removes markdown code formatting and extracts the content of the body tag if the model returns a complete HTML page. This eliminates the need to add formatting restrictions to the prompt.

Before writing to the information block, the response undergoes comprehensive validation: length in characters and words, format (text, HTML, or JSON), language based on the proportion of alphabet letters, and the presence of forbidden or mandatory phrases. A rejected response never enters the information block: the step is either skipped or the entire pipeline is stopped, depending on the selected setting.

The module ensures fault tolerance by automatically retrying requests with increasing delays when temporary provider errors occur. If an invalid access key is detected or funds are exhausted, the queue stops immediately, preventing thousands of identical failures and unexpected charges.

Comprehensive control and reporting features are included. Each step can be configured to require moderation, turning results into pending proposals. Editors can view the current field value alongside the suggestion, then accept, edit in place, or reject with a reason—either individually or in batches. Reverting applied changes to the previous value remains possible even after a week without needing to restore a database backup.

A dashboard and run chronology provide full visibility. For each card, the system displays the prompt sent to the model, the values inserted into each field, model responses before and after processing, validation results, and final recorded data. Step settings are captured as snapshots at the moment of execution, ensuring the history reflects exactly how the card was processed.

Token usage and cost tracking are recorded per step and aggregated by run. The system logs input and output token counts, the actual cost charged by the provider, and generation time. This allows users to identify expensive steps, such as those involving web search.

This module automates text content generation for information blocks in the 1C-Bitrix system. It supports any platform edition that includes the Information Blocks module: Standard and higher, Small Business, Business, as well as 1C-Bitrix: Site Management editions containing information blocks. PHP version 8.1 or newer is required. A properly configured task scheduler (cron agents) on the site side is necessary to maintain the background queue.

Generation requires an account and access key from a neural network provider: Polza.ai, OpenRouter, or a custom OpenAI-compatible endpoint. Token costs are paid directly to the provider according to their tariffs; the module does not charge a fee for generation. The server must have outgoing access to the provider's API.

The module operates exclusively with text. The current version implements only one step type: text generation. Images are not created, and generation results are not written to image or file fields. Only information blocks are supported; Highload blocks are not yet supported but are planned. Information block elements and sections are processed correctly. Catalog products are treated as standard information block elements, meaning the module cannot access data such as price, stock quantity, or unit of measure.

The generation process follows a linear pipeline. Steps execute strictly in sequence without branching or conditional logic between them. All filter conditions are combined using logical AND. Separate profiles must be created to handle different sets of conditions.

The module operates in the background with a fixed speed: each agent launch processes three cards with a one-minute interval, resulting in approximately three cards per minute per profile. These values are not configurable in the current version. Mass generation for large catalogs should be planned for extended periods.

Model parameters cannot be adjusted. The interface lacks settings for temperature, token limits, or similar options; the result is controlled by the prompt, model selection, and answer verification. Streaming output, function calling, and sending images as input are not supported.

An error at any step stops the pipeline for that specific card. No data is written to the fields, including results from previously successful steps.

Reporting is itemized. History is maintained per card and run. There are no summary reports for a period, charts, file export, or automatic history cleanup by date.

No notifications are available. The module does not send emails or messages to Bitrix24 regarding queue stops or new moderation proposals. Status is visible only on the dashboard.

Permissions do not include a separate role for confirming results. Actions are available to anyone with access to the module settings.

The module automates the filling of fields, properties, and SEO parameters for information blocks using artificial intelligence. The demonstration mode is limited to 500 generated records per installation; further generation stops until a license is activated.

A personal neural network API key is required. The module acts as a bridge between Bitrix and the provider, operating with your key and balance. This approach is more transparent than all-inclusive subscriptions: you pay for tokens at the provider's rates and can view exactly where funds are spent at any time.

Payment can be made in rubles via the Russian aggregator Polza.ai, which provides access to models such as GPT, Claude, Gemini, DeepSeek, and others without the need for a VPN or foreign bank card. Product card data is sent only to the provider whose key is entered in the profile settings. There are no intermediate servers; statistics and run history are stored in the module's tables on your site. If data cannot leave the perimeter, a local model or a custom OpenAI-compatible gateway can be connected.

The module does not affect descriptions that were previously written manually.

The module preserves manual content: the 'field not filled' filter skips existing data, while the 'fill only empty field' mode protects data at the step level. Any change can be rolled back to the previous value.

The cost of generating content for the entire catalog is calculated before launch. A test run on a single card shows the actual cost per step, and the filter displays the exact number of items in the selection. Multiplying these values yields the total budget. Resource consumption is visible in real time for each run.

Text quality control is ensured by three protection levels. Preliminary checks filter out responses that are too short, use the wrong language, have an incorrect format, or contain prohibited phrases before saving to the infoblock. The moderation stage formats the result into a proposal that an editor can accept, edit, or reject. The rollback function restores the previous field value if the decision is changed later.

The module supports generating SEO texts for catalog sections. Sections are treated as full-fledged entities: the system reads their fields, descriptions, and user-defined fields, and writes the title, symbolic code, description, and SEO templates.

The module correctly processes cards imported from 1C.

The module automatically queues new or modified information block elements for processing if they pass the profile filter, provided that event capture is enabled. Background processing ensures that import speed is not affected.

Deep expertise in writing prompts is not required. Fields can be inserted into the prompt by clicking from a list. The 'Improve Prompt' button allows rewriting the text for greater accuracy or generating it based on a task description. Successful variants are saved to a collection for reuse.

The module can handle a catalog of ten thousand items, though not instantly. Processing occurs in batches of approximately three cards per minute per profile. The queue maintains progress and continues operation even if the admin panel is closed or the server is restarted. To meet tight deadlines, it is advisable to split the catalog across multiple profiles.

Implementation does not require a programmer. All settings, including provider connection, filters, and pipeline steps, are configured in the admin panel. A programmer is only needed to create a non-standard gateway to a custom model.

General iT can handle the implementation of “AI Content Generation: filling fields, properties, and SEO for information blocks using AI”, from installation to configuration and compatibility checks.