Automated product descriptions
From manual enrichment to a guarded multi-source pipeline — with a quality gate before publish.
- Client
Medizinfuchs GmbHPharmacy price-comparison platform
At a glance
- Situation
- Product descriptions were written by hand, one product at a time.
- Goal
- Generate them automatically — without softening the duty of care on health content.
- Solution
- A multi-source pipeline with a fixed schema, guardrails, and a quality gate before publish.
- Result
- ~70% less manual content creation, and incomplete products never go live.
Manual work did not scale
Product data enrichment and descriptions were created by hand, product by product.
Before
- Open and compare sources per product
- Write attributes and copy manually
- Check mandatory legal notices one by one
After
- Ingest 5–12 partner sources automatically
- Extract, merge, and guard structured content
- Publish only descriptions that pass the gate
What “done” had to mean
Automation was only an option if it did not dilute the duty of care. So the target and the limits were fixed before the first line of code.
Success criteria
- Descriptions without manual research per product
- Several partner sources condensed into one consistent text
- A fixed section structure instead of free-form prose
- Scalable across the whole catalogue, not just single cases
Non-negotiable
- No facts that are not in a source
- Mandatory information complete — or no publish
- Contradictions made visible, never averaged away
- Every rejection logged with a traceable reason
Same pattern on your side?Book a short call
From PZN to published description
Every step has a clear job. Products missing required information never go live.
Every product is identified by its PZN — the entry point for the full pipeline.
For each PZN, product pages from 5 to 12 partner online pharmacies are used as sources — including Aponeo, apolux, Pharmeo, Shop-Apotheke, and DocMorris.
Product pages are scraped at scale and normalized into a consistent format.
An LLM maps each source into a fixed schema: attributes, safety, benefits, FAQs, and linked documents.
Collect → categorize → reconcile → prioritize by source trust → synthesize. Hard contradictions follow source priority; equal trust keeps both values under “Unassigned”.
No invented facts, no interpretation, no generalization, no averaging of conflicts. Mandatory notices for dietary supplements are added automatically.
Products missing required sections or information are not published. The reason is logged for review.
Rejected products land in review with a clear reason — nothing goes live silently.
Only approved products appear with up to 14 sections in a fixed order — empty sections are skipped.
Click or tap for details.
5–12 sources, one structured description
Each product pulls in five to twelve partner sources. The pipeline categorizes, reconciles, and prioritizes by source trust — instead of blending prose.
- Collect
- Categorize
- Reconcile
- Prioritize
- Synthesize
Health content under YMYL rules
The pipeline may merge facts — not invent or interpret them.
Allowed
- Extract and assign facts from sources
- Merge complementary details
- Add mandatory dietary-supplement notices automatically
- Resolve contradictions by source priority or show them transparently
Not allowed
- Invent facts
- Interpret or generalize sources
- “Resolve” conflicts by averaging
- Publish incomplete products
Each task on the cheapest model that still holds quality.
No single model for the whole pipeline. Steps are routed separately, models are compared, and expensive capacity is used only where it actually moves the output.
- 01
Route by task
Structured extraction mainly needs schema fidelity — cheaper models are enough there. Synthesis and guardrails touch release and liability; stronger models run there.
- 02
Compare, don’t guess
GPT, Gemini, and Claude are measured per task against quality and cost. Routing follows the numbers — which model lands on which step is a result, not a preference.
- 03
Pipeline built for cost
Cheap steps run often and in parallel; expensive ones only when validation or YMYL risk requires them. Run cost stays low without watering down the duty of care.
Up to 14 sections — fixed order
Only sections with content are emitted. Empty fields are skipped.
- Main description
- Properties
- Use
- Recommended intake
- Instructions for useNo data — skipped
- Storage
- Ingredients
- Composition (table)
Example: composition (excerpt) Substance Amount Vitamin C 80 mg Zinc 10 mg Biotin 50 µg - Allergy notices
- Nutrition tableNo data — skipped
- Size chartNo data — skipped
- Warnings
- Manufacturer data
- UnassignedNo data — skipped
Less manual work — same duty of care
The pipeline handles scraping, extraction, synthesis, and quality control. Release stays rule-based — with no invented facts.
~70%
less manual content creation
- 5–12
- partner sources per product, merged automatically
- 14
- sections at most, in fixed order — empty ones skipped
What deliberately stayed the same
- Nothing goes live without complete mandatory information
- No invented facts — only what a source supports
- Rejected products land in review with a reason
Same pattern, different industry
The pipeline was built for pharmacy products — the pattern was not. It applies wherever people condense data from many sources into text by hand.
Many sources, one record
Your people open several supplier, manufacturer, or partner sources per record and reconcile them in their heads.
Content with liability attached
Mandatory notices, legal text, or safety information have to be right — “mostly correct” does not clear the bar.
Throughput scales with headcount
More products, markets, or languages currently means more people doing the same work.
If two of these fit, a call is worth it. Twenty minutes is enough to tell whether the pattern carries.