Skip to content

SYSTEM Cited by 1 source

Zalando Prompt Generator

What it is

Prompt Generator is the LLM orchestration service inside Zalando's Content Creation Copilot. It sits between the Content Creation Tool + Article Masterdata on the inbound side and the LLM backend (OpenAI GPT-4 Turbo → GPT-4o) on the outbound side. It is the load-bearing component of the copilot — the thing that turns a schema + an image into a prompt and a model response into structured attribute suggestions (Source: sources/2024-09-17-zalando-content-creation-copilot-ai-assisted-product-onboarding).

Responsibilities

1. Prompt materialisation

Given an article's attribute set from Article Masterdata and a set of image URLs from the Content Creation Tool, construct the LLM prompt:

  • Include English labels for attribute codes.
  • Include allowed-value vocabularies for each attribute.
  • Include custom per-attribute guidelines (e.g. distinguishing V-neck from Low cut V-neck collar) crafted by content-creation experts. "We created custom guidelines as part of the prompt for complex product attributes which gave additional hints."
  • Reference the selected input image(s).

2. Code ↔ English translation (both directions)

Translate Zalando's opaque attribute codes (assortment_type_7312) to English (Petite) on the way into GPT, then translate the English response back to codes on the way out. This is the concepts/opaque-attribute-code-translation-layer pattern: the LLM speaks natural language, the catalog speaks identifiers, the Prompt Generator is the bidirectional shim.

From the post: "We built a translation layer that converts OpenAI output into information directly usable by Zalando and discards the part that is not relevant."

3. Category-aware attribute filtering

Apply the concepts/category-attribute-relevance-mapping filter before prompting: if an attribute shouldn't be filled for this category (per internal guidelines) or if GPT accuracy on this attribute × category pair was empirically poor, remove it from the prompt entirely. Scope-reduction as a quality lever.

4. Image selection

Pick the highest-signal image from the article's available set. Empirically (concepts/input-image-selection-tradeoff): "product-only front images delivering the best results, followed closely by front images featuring the products being worn by the model." Other image types (detail shots, back views, lifestyle) produced worse suggestions and cost tokens, so they're deprioritised.

5. Output parsing + discard

Parse the LLM response, map English values back to codes, and discard irrelevant output (hallucinated attributes, malformed JSON fields) before returning to the Content Creation Tool.

Model-agnostic dispatch (aggregator role)

The post frames the copilot as an aggregator service: the Prompt Generator is the integration point where new backends (brand data dumps, partner contributions, fine-tuned models) slot in without changing the Content Creation Tool contract. During development, GPT-4 Turbo was swapped out for GPT-4o with no downstream contract changes — evidence the aggregator framing actually works.

Design trade-offs

  • Schema-driven over template-driven. Prompts are assembled from Article Masterdata at call time, not stored as static templates. Advantage: new attributes immediately available. Trade-off: prompt drift across attribute types is harder to audit — there's no single "the prompt" to review, just the generator logic and the per-attribute guidelines.
  • Translation layer is Zalando-internal, not LLM-internal. An alternative would be to fine-tune a model on Zalando's attribute codes directly. The Prompt Generator approach preserves model portability at the cost of always paying the translation overhead on every call. Given the model swap (GPT-4 Turbo → GPT-4o) disclosed in the post, the portability premium appears to have paid off at least once.

Seen in

Last updated · 501 distilled / 1,218 read