Behind the scenes of the AI workflow: How multiple LLMs are working together to deliver future-oriented product data

A look behind the scenes of AI-supported product data enrichment

The management and optimization of product information has fundamentally changed in recent years. While traditional PIM (Product Information Management) systems rely on manual inputs and rule-based automation, Ainavio uses an innovative multi-LLM (Large Language Models) concept to intelligently enrich product data. What does multi-LLM concept mean and how exactly do these models work together? In this article, we take a look at Ainavio's advanced AI workflows that combine various specialized LLMs to achieve unparalleled data quality.

Why a single LLM isn't enough

Many companies are already using AI to automate product data processes. But an individual LLM quickly reaches its limits: While one model is ideal for text creation (such as OpenAI language model), another one shines in attribute recognition (Anthropic Claude) or yet another for extracting texts from PDFs (Google Gemini 2.0 Flash). Ainavio therefore goes one step further and combines several specialized LLMs in a multi-stage process.

In addition, the individual providers in turn have various stages of expansion of their language models, with different prices. Here, too, ainavio uses various models depending on requirements to optimally manage costs and data quality.

By dividing tasks in this way, different aspects of product data can be processed and optimized in parallel. The result is a seamless, automated workflow that not only significantly improves speed, but also quality and consistency.

Ainavio's multi-LLM approach: synergies between specialized AI models

Ainavio uses specialized AI models that perform various tasks within the PIM process. The combination of these models ensures optimal data quality and process automation. Here is an example workflow:

1. Data extraction and preprocessing --> Google Gemini Flash 2.0 language model

• AI-powered analysis of incoming data from supplier feeds, datasheets, and web sources.
• Discover structured and unstructured data.
• Remove duplicates and correct inconsistent information.

2. Attribute generation and completion --> Anthropic Claude language model

• A specialized LLM extracts relevant attributes from texts, images, or tables.
• If certain information is missing, the AI automatically supplements the data with pattern recognition.

3. Automatic text creation and optimization --> OpenAI language model

• Another LLM generates SEO-optimized product descriptions based on the acquired attributes.
• Style, tonality and keywords are adapted to the respective target group.
• Automatic spelling and grammar checker.

4. Multilingual localization and metric conversion --> depending on the language DeepL

• An LLM specialized in languages translates product data into various markets.
• Conversion of measurement units (such as inches to centimeters) and currency conversions depending on the target region.

5. Quality control and validation --> OpenAI language model

• A final LLM checks whether all product information is accurate, complete and in line with the market.
• Manual intervention by the team is only required in special cases.

Expansion options: integration of external services

In addition to these specialized LLMs, other services can be integrated into the workflow to further improve data quality. For example, it enables Perplexity AIto carry out AI-based web research to add missing or incomplete information. This combination of LLMs and external tools makes the workflow even more powerful and ensures a complete, precise and up-to-date database.

The combination of several specialized LLMs in a structured AI workflow is the key to outstanding product data quality. Each model performs a specific task and helps to ensure that the data is not only precise, but also comprehensive, SEO-optimized and adapted to specific markets.

Precise control through workflows and prompts

These AI processes are controlled by a combination of automated workflows and individually defined prompts:

• Automated control through Ainavio workflows: Based on the current data situation, the system decides in real time which LLM and which processes are activated. If a quality check shows that the results do not meet the requirements, readjustment is carried out automatically.
• Individual definition through prompts: Rules, requirements and dependencies such as attributes, texts or formats are controlled by precisely formulated prompts. This ensures that AI delivers consistent and brand-compliant results. These prompts are stored individually in the ainavio PIM system and can be changed by any user.

Example 1: Optimizing product descriptions

For the “description text” attribute, the prompt could be as follows:

• Create an SEO-optimized product description for the web shop
• 200 - 300 words, continuous text
• Mention the most important USPs, but do not repeat any from the bullet points (another attribute)

This targeted prompt ensures that the generated text is not only formulated attractively, but is also optimized for search engines. A dependency is also defined directly, no repetitions to the other “Bulletpoints” attribute.

Example 2: Defining a dependency between two attributes

For the “Drinking Temperature” attribute, the prompt could read as follows: “Add the recommended drinking temperature based on the wine type. White wine: 10—12°C, red wine: 16—18°C, rosé: 10—12°C, sparkling wine & secco: 6—8°C.”

This targeted prompt ensures that the AI automatically adds the appropriate drinking temperature for each type of wine, resulting in a uniform and logical product description.

Example 3: Raw data conversion — correct incorrect format

For the “Package weight” attribute, the prompt could read: “Convert the package weight from grams to kilograms. Example: 160 g → 0.160 kg.”

This prompt ensures that supplier data, which is available in various formats, is automatically converted and standardized to ensure consistent data output.

Example 4: Extraction from product image

For the “vegan” attribute, the prompt could be formulated as follows: “If a 'vegan' icon is visible on the packaging, the attribute is automatically filled in with 'yes/true. '”

With this prompt, the AI can extract visual information from product images and automatically add relevant attributes, which significantly reduces the manual maintenance of such information.

Another advantage of this approach: Machine learning continuously improves prompts so that the AI delivers ever more precise and relevant results.

These examples only show a small selection of possible areas of application. The customization options are almost unlimited and can be individually tailored to your requirements. We would be happy to demonstrate this live in our system with your own product data.

The business case: Why companies benefit from multi-LLM

Companies that rely on a multi-LLM strategy for their product data benefit from several decisive advantages:

• Time savings: Automated data processing drastically reduces manual effort.
• Higher data quality: AI-supported validation significantly reduces the error rate.
• Faster time-to-market: New products can be released on various platforms in record time.
• Flexibility for global markets: Automatic localization and metric conversion enable international expansion without additional effort.

‍