How to output results in CSV format instead of markdown from Palm API?

Question

I'm new to using the PalmML API and I"m trying to generate results that return in CSV format. No matter what I try, I am getting markdown format. Here is example prompt.

              const result = await client.generateText({
                model:MODEL_NAME,
                prompt: {
                    text: 'Generate a list of top 10 websites and return results in comma separated value format, not markdown format. Put field names in first row.',
                    format: 'csv',
                    delimiter: ',',
                    quote: '"', 
                  },
                 
               });

The output is always in markdown format like following:

| Rank | Website | Alexa Global Rank |
|---|---|---|
| 1 | Google | 1 |
| 2 | YouTube | 2 |
| 3 | Facebook | 3 |
| 4 | Baidu | 4 |
| 5 | Wikipedia | 5 |
| 6 | Amazon | 6 |
| 7 | Tencent | 7 |
| 8 | Twitter | 8 |
| 9 | Instagram | 9 |
| 10 | Reddit | 10 |

Is there a way to have it output CSV format? Or Do I need to convert markdown to CSV on my own?

Answer 1

Firstly you can remove these fields as they aren't going anywhere (see the API docs):

prompt: {
  format: 'csv',
  delimiter: ',',
  quote: '"',
},

There are a few approaches that will probably help. As LLMs are just auto-complete services, you can try giving it the start of the CSV example in a zero shot request (I would also remove not markdown format to avoid including references to markdown format), e.g.

Zero-shot prompt:

Generate a list of top 10 websites and return results in comma separated value format.

websites.csv:
rank,website,url

You can often do better with a few-shot prompt by giving examples.

Few-shot prompt (1-shot really):

Task: Top-5 ice cream flavours in CSV format.
icecream.csv:
rank,flavour,colour
1,chocolate,brown
2,vanilla,white
3,strawberry,pink
4,banana,yellow
5,mint choc-chip,green

Task: Top-10 websites in CSV format
websites.csv:
rank,name,url

---

Another approach is to get the model to come up with results in stages.

Prompt 1:

The top 10 websites by traffic are

Response:

The top 10 websites by traffic are:
1. Google
2. YouTube
...

Prompt 2:

The top 10 websites by traffic are:
1. Google
2. YouTube
...

The above list in comma-separated format is:
rank,website

You can even do this dynamically.

Prompt:

The top websites, in CSV format are:

Try to evaluate the results, if it can't be coerced into CSV then append the prompt with a correction stanza.

The top websites, in CSV format are:
| Rank | Website |
|---|---|
| 1 | Google |
...

That was not CSV, in comma-separated format, it is:

This last approach is a bit better suited to the generateMessage API but you can make it work in text too.