Automated Product Attribute Extraction and Title Standardization Using Agentic AI

delivery_hero · E-commerce · 2025 · source Agentic: yes · promotional bias: high · summary-only (no fetched source) · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Quick-commerce product catalog management

Delivery Hero Quick Commerce needed to keep product attributes and titles accurate and standardized across large regional catalogs without relying on slow, costly, error-prone manual verification.

Constraints:

catalogs span multiple geographical regions and platforms
vendor-provided product titles can be incomplete or inconsistent
manual attribute verification does not scale
the production workflow needs predictable cost, latency, and debuggability
uncertain predictions still need human quality review

Architecture

Delivery Hero uses a predefined two-step LLM agent workflow for catalog enrichment. The first LLM component receives product titles and images and extracts 22 predefined attributes such as brand, flavor, and volume. The second component uses the extracted attributes to generate standardized product titles in Delivery Hero QC's format. The team chose predefined sequencing over dynamic orchestration for lower fixed cost, consistent latency, predictability, and easier debugging. Prompt engineering reduces ambiguity and token cost, while a teacher/student distillation strategy uses a strong teacher model to create standardized-title data for a smaller student model. Confidence scores from output logits flag uncertain outputs for human review.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  vendor-product-input[Vendor product input] --> attribute-extraction-component[Attribute extraction component]
  prompt-optimization-layer[Prompt optimization layer] --> attribute-extraction-component
  attribute-extraction-component --> title-generation-component[Title generation component]
  prompt-optimization-layer --> title-generation-component
  teacher-model[Teacher model] --> student-model[Student model]
  student-model --> title-generation-component
  title-generation-component --> confidence-scoring-gate[Confidence scoring gate]
  confidence-scoring-gate --> human-review-queue[Human review queue]

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Vendor product input`	Provides product titles and images from vendors.	`Attribute extraction component`
`Attribute extraction component`	Extracts 22 predefined product attributes from titles and images.	`Vendor product input`, `Title generation component`
`Title generation component`	Generates standardized product titles from extracted attributes.	`Attribute extraction component`, `Confidence scoring gate`
`Prompt optimization layer`	Keeps prompts concise and unambiguous to reduce token cost, latency, and task misunderstanding.	`Attribute extraction component`, `Title generation component`
`Teacher model`	Creates high-quality standardized title examples using detailed prompts.	`Student model`
`Student model`	Replicates the title-formatting behavior with shorter prompts and lower production cost.	`Title generation component`
`Confidence scoring gate`	Converts logits into probability scores and routes low-confidence outputs for review.	`Title generation component`, `Human review queue`
`Human review queue`	Provides manual verification for edge cases and uncertain predictions.	`Confidence scoring gate`

Data flow

Vendor product titles and images flow into a multimodal LLM attribute extractor. Extracted attributes then flow into a title-generation LLM that produces standardized product titles. Prompt engineering and teacher/student fine-tuning reduce the cost and latency of title generation, while confidence scoring routes low-confidence attributes or titles to human review.

3 supporting quotes

“The production system consists of two core LLM components orchestrated sequentially.”summary, stated

“The first component handles attribute extraction, receiving vendor product titles and images as inputs and extracting a comprehensive set of 22 predefined attribute types including brand, flavor, volume, and other relevant characteristics.”summary, stated

“The second component performs title generation, taking the extracted attributes and generating a new product title”summary, stated

Why this architecture

The two-step predefined agent matches the business task: extract attributes first, then generate a title from those attributes. Delivery Hero prioritized operational predictability over maximum flexibility because catalog enrichment is a core production process. Teacher/student distillation moves formatting expertise from expensive prompts into a smaller model, and confidence gating keeps quality review focused on uncertain outputs.

4 supporting quotes

“The two-step task structure—first extract attributes, then use those attributes to generate a standardized title—naturally mapped”summary, stated

“Their priorities centered on efficiency, cost-effectiveness, predictability, and debuggability for a core cataloging process.”summary, stated

“This knowledge transfer method significantly reduced operational costs and latency for the title generation component”summary, stated

“Outputs falling below a predefined confidence threshold are automatically flagged and routed for human review.”summary, stated

Models used: Large Language Models · multimodal Large Language Models · GPT-4o · GPT-4o-mini · teacher LLM · student LLM

Scale

The system extracts 22 predefined product attributes.
“The first component handles attribute extraction, receiving vendor product titles and images as inputs and extracting a comprehensive set of 22 predefined attribute types including brand, flavor, volume, and other relevant characteristics.”summary, stated
Delivery Hero Quick Commerce manages catalogs across multiple geographical regions and platforms.
“Delivery Hero Quick Commerce operates in the fast-paced quick commerce sector, managing extensive product catalogs spanning multiple geographical regions and platforms.”summary, stated

Alternatives considered (roads not taken)

Manual product-attribute verification

Rejected because: Manual verification was time-consuming, expensive, error-prone, and did not scale with catalog needs.

“The manual verification was not only time-consuming and expensive but also inherently prone to human error, creating inconsistencies that cascaded through their operations affecting inventory management, search relevance, and ultimately customer satisfaction.”summary, stated

Single-turn LLM query

Rejected because: A single-turn query was suitable for simpler tasks but not enough for the two-step extraction and title-generation workflow.

“single-turn LLM queries suitable for simpler tasks like summarization but insufficient for their multi-step requirements”summary, stated

Dynamically orchestrated agents

Rejected because: They provide flexibility and self-correction but raise cost, latency, unpredictability, and debugging difficulty.

“Dynamically orchestrated agents, conversely, provide higher flexibility with dynamic path adjustment and self-correction”summary, stated

“but at the cost of higher and variable expenses, increased latency, lower predictability, and more challenging debugging”summary, stated

Patterns

`prompt-chaining` (orchestration)

Delivery Hero chains an attribute-extraction LLM step into a title-generation LLM step.

“The production system consists of two core LLM components orchestrated sequentially.”summary, stated

`model-distillation` (model-operations)

A GPT-4o teacher creates standardized title examples used to fine-tune a smaller GPT-4o-mini student.

“a powerful "teacher" LLM such as GPT-4o was used with detailed prompts containing many examples”summary, stated

“this dataset was used to fine-tune a smaller, more efficient "student" LLM such as GPT-4o-mini.”summary, stated

`confidence-thresholding` (governance)

Output-logit probabilities determine which extracted attributes or titles need manual verification.

“The implementation processes output logits from the LLMs and converts them into probability scores”summary, stated

`human-in-the-loop-gate` (governance)

Low-confidence outputs are routed to human reviewers to maintain catalog quality.

“Outputs falling below a predefined confidence threshold are automatically flagged and routed for human review.”summary, stated

`cost-aware-routing` (operations)

The predefined agent path and student model reduce variable LLM calls, prompt length, latency, and cost.

“predefined agents offer lower and fixed costs with fewer LLM calls, lower and consistent latency”summary, stated

Tradeoffs

Use predefined agents

Chose fixed two-step LLM orchestration over dynamically orchestrated agents.

Rationale: The catalog workflow valued predictability, debuggability, fixed cost, and consistent latency more than flexible self-directed execution.

“predefined agents offer lower and fixed costs with fewer LLM calls, lower and consistent latency, higher predictability”summary, stated

Distill title generation into a smaller model

Chose fine-tuned GPT-4o-mini student over teacher model with long prompts at inference time.

Rationale: The student could keep teacher-like output quality while reducing prompt length, cost, and latency.

“the fine-tuned student model could replicate the teacher's output quality but with much shorter and more efficient prompts.”summary, stated

Use confidence scoring for review routing

Chose automatic processing with low-confidence human review over manual review of every output.

Rationale: Confidence thresholds balance automation benefits with quality control for uncertain edge cases.

“maintaining quality standards while still achieving significant automation benefits.”summary, stated

Gotchas (what bit them)

Vendor titles can omit attributes or use inconsistent formats.

Root cause: Different vendors structure product titles differently and may leave out crucial details.

Resolution: Use multimodal LLMs to extract attributes from title and image inputs before generating a standardized title.

“vendor-provided product titles often lacked crucial attributes or failed to adhere to Delivery Hero's standardized format”summary, stated

“This required a solution that could both extract attributes and restructure titles into a consistent format.”summary, stated

Long example-heavy prompts increase cost and latency.

Root cause: The title-generation task initially needed many examples inside the prompt to produce the standardized format.

Resolution: Use teacher-generated examples to fine-tune a smaller student model that needs shorter prompts.

“this approach led to lengthy prompts that increased both processing time and costs.”summary, stated

“after a few training cycles, the fine-tuned student model could replicate the teacher's output quality”summary, stated

The public case study omits important production-monitoring details.

Root cause: It emphasizes success without explaining drift detection, threshold selection, or reviewer feedback loops in depth.

Resolution: Treat the architecture as useful but note that production operators would still need monitoring and feedback-loop design.

“the case study doesn't discuss monitoring and observability practices”summary, stated

“or how they handle updates to the system as product catalogs, vendor behaviors, or business requirements evolve.”summary, stated

Evaluation

The summary reports qualitative improvements and says student-model evaluations replicated teacher output quality, but it does not provide specific accuracy, latency, confidence-threshold, or reviewer-load metrics.

Metrics: claimed efficiency improvement · claimed accuracy improvement · claimed data-quality improvement · claimed customer-satisfaction improvement · student model replication of teacher quality, unspecified metric

“The claimed improvements in efficiency, accuracy, data quality, and customer satisfaction are presented without specific metrics”summary, stated

“Evaluations demonstrated that after a few training cycles, the fine-tuned student model could replicate the teacher's output quality”summary, stated

Outcomes

The system automates extraction of 22 product attributes and standardized title generation.
The public summary claims improvements in efficiency, accuracy, data quality, and customer satisfaction, but without quantitative metrics.
Teacher/student distillation reportedly reduced title-generation cost and latency while preserving quality.

The case is summary-only and describes multi-region/multi-platform quick-commerce catalogs rather than giving exact catalog size or production throughput.