Extracting Information

Information Extraction Prompt Examples

• Names and contacts: 'Extract all names and email addresses from this document. Format: Name | Email, one per line.'
• Key numbers: 'Extract all financial figures from this report. Include: the metric name, the value, and the context (what it refers to).'
• Dates and deadlines: 'List all dates and deadlines mentioned in this document. Format: Date | Event | Responsible party.'
• Action items: 'Extract all action items from these meeting notes. Format: Action | Owner | Deadline.'
• Key claims: 'List the 5 most important claims made in this research summary. Include the page or section reference if available.'

Extraction Prompt Best Practices

# Information extraction best practices

extraction_prompt = """
Extract [WHAT TO EXTRACT] from this document.

Format the output as:
[SPECIFY THE EXACT FORMAT, e.g. Column 1 | Column 2 | Column 3]

If [ITEM] is not found in the document, write 'Not mentioned'.
Do not add information that is not in the document.

[PASTE DOCUMENT]
"""

# Key best practices:
# 1. Specify exactly what to extract
# 2. Define the output format precisely
# 3. Tell AI what to do if the item is missing
# 4. Include 'do not add information not in the document'
# 5. Always verify the extraction against the original

print("Extraction prompts need precise instructions on what to find.")
print("Always verify extracted data against the source document.")

Key Takeaways

• Extraction prompts should specify exactly what to look for and the output format
• Tell AI what to do when information is not found in the document
• Instruct AI not to add information that is not in the source
• Always verify extracted data against the original document
• Extraction is most reliable for structured information like names, dates, and numbers