I just used Rytr to generate property descriptions from raw listing data. My ETL pipeline pulls MLS feeds into BigQuery, but the descriptions are... not great. Rytr's "Formal" tone actually works for high-end listings.
I set up a simple Python script in my pipeline to call Rytr's API after the raw data lands. Here's the basic transform step:
```python
def enhance_listing(row):
prompt = f"Write a formal real estate description for a {row['bedrooms']} bedroom property..."
# Call Rytr API here
return polished_description
```
**Observations so far:**
* It's consistent with terminology (sq ft, amenities, "spacious")
* Sometimes repeats phrases if the input data is sparse
* Need to watch the token usage with bulk processing
Has anyone else tried using it for structured data enhancement? I'm curious about:
* Batch processing strategies
* Cost vs. quality for thousands of listings
* Whether I should do this pre-load or post-load in the warehouse
Neat use case, but I'd worry about blindly piping this into production. Rytr's API might be stable, but you're adding a live external dependency to your ETL. Ever think about running a local model in a container for this step? You could bake something like a fine-tuned T5 or even a small GPT-NeoX variant into a pipeline stage.
For batch jobs, the cost/quality ratio gets weird at scale. I'd test a sample through something like GPT-3.5 Turbo via API and compare it to your Rytr output. Might be cheaper per listing in bulk, and you'd have more control over repetition.
On your last question: do it pre-load, but cache the hell out of the results. You don't want to re-generate descriptions every time you rebuild a mart. Store the polished text as a new column right in the warehouse table. Keeps the raw data intact and lets you version the enhancement logic separately.
My sandbox is bigger than yours.