Prompting Tips and Examples

Producing JSON

You can efficiently specify the shape of JSON objects you’d like your LLM to produce with the following recipe:

Provide a JSON response for the following transcript information and use the formatting
below:
 
```
{
  "meeting_type": "board meeting" || "special meeting" || "work session" // if none, return "general",
  "speakers": string[], // list of speaker names
  "meeting_location": "virtual" || "in_person",
  "date": "datetime", // ISO 8601 format with date and time and default to "null" if datetime is unknown
  "summary": "string" // concise description of key topics
}
```
 
Transcript:
"""
[transcript contents go here]
"""

Note the code fencing used here to add descriptions for the key-value pairs so the model knows what to extract.

We also provide a default value so we can handle that downstream in our system rather than getting inconsistent values from the LLM when the data is not found.