Using a loop, we pass the keywords in batches and ask in JSON format. Here, the prompt (the input with the instructions we insert) is important. I leave you an article that talks about it and puts on the table 26 proven rules to improve your prompts (you can also find some tips on the official OpenAI page ).
It should be not that in each interaction of
The loop there are three ways that the loop can fail. Because I don’t classify all the words,
Cecause it does not classify the keywords in the groups of the general categories pass in the input,
Because it fails (usually because it returns the JSON in an incorrect format or because it cuts the JSON).
In these cases, a margin of retries
Must be allow for the same iteration. In each retry, the keywords must be pass in a different order, since this substantially improves the results. If after these attempts there are still problems, the unclassifi or poorly classifi keywords will be sav ig database to be us in the next step. Therefore, at the end of this second step we must have two lists, the list of well-classifi keywords and the list of poorly classifi keywords.
Step 3. The play-off:
In step 3 we will try to re-fetch the poorly rank or unrank keywords with another loop and shuffling the keywords again. If in this step we do not get a how to avoid food waste in catering? good ranking for any of the keywords, they will be mark as not being able to rank (it tends to not reach 0.5% of the total). In our test, we launch 430,000 keywords in 9 days. Just under 2000 of them fail to rank:
Using 3.5 with 4 refinement
This is a somewhat more expensive approach, but its usa data results are, in principle, more accurate (and without the ne to invent the categories in advance). Let’s see how the steps differ from the first methodology.