Summaries and Summarization
Please see the general documentation on Summarization using generative AI
Overview of Structured (JSON) Summaries
Structured (JSON) Summaries are summaries whose output is formatted as valid JSON. They enable easier parsing, integration with third-party systems, and more straightforward data processing than free-form text summaries.
{
"CompanyMentioned": "ExampleCorp",
"MentionContext": "Customer mentions CompetitorCorp as a competitor",
"MentionSentiment": "Positive"
}More details on JSON structured format
When to Use JSON Summaries
• You plan to import the summary data into analytics or reporting tools (e.g., splitting each JSON key into its own column in a spreadsheet).
• You want to ensure a consistent, easily readable (machine-readable) format.
• You need to enforce specific fields or data points in every response.
• You need to extract/cover more than one piece of information from the conversation.
Enabling JSON Output
1 - Enable Force JSON mode in GenAI summary
This setting ensures the output will always appear in a valid JSON structure.
2 - Add instructions to the Summary Request / Prompt
Specify which JSON keys should appear in the response and how values will be filled.
Example
Please output your answer in JSON format with the following keys: CompanyMentioned, MentionContext, MentionSentiment.
When the AI processes a conversation or text according to these instructions, it will respond with something like:
{
"CompanyMentioned": "ExampleCorp",
"MentionContext": "Customer mentions CompetitorCorp as a competitor",
"MentionSentiment": "Positive"
}Structuring the Summary Request / Prompt
You can structure your prompt in multiple ways:
Option 1:
In the summary request, explicitly state the keys and how you want them populated:
Analyze conversation and answer next questions:
- Question #1 ...
- Question #2 ...
- Question #3 ...
Please provide the summary as a JSON object with the following keys: Key-1, Key-2, Key-3
Example:
- Record which competitor’s name the client mentioned.
- Describe the context in which this competitor is mentioned.
- Determine the sentiment (positive, negative, neutral).
Please provide the answer in JSON format with the keys: CompanyMentioned, MentionContext, MentionSentiment.
Option 2:
Give brief instructions stating all keys in one place. For example:
Analyze conversation and provide the answer in JSON format:
- 'Key-1': Question #1 ...
- 'Key-2': Question #2 ...
- 'Key-3': Question #3 ...
...
Example:
Analyze conversation and provide the answer in JSON format:
- 'CompanyMentioned': Question #1 ...
- 'MentionContext': Question #2 ...
- 'MentionSentiment': Question #3 ...
...
Both options should lead to a stable JSON response:
{
"CompanyMentioned": "ExampleCorp",
"MentionContext": "Customer mentions CompetitorCorp as a competitor",
"MentionSentiment": "Positive"
}
Benefits of Structured (JSON) Summaries
• Consistency: Every summary follows the same layout, making data extraction predictable.
• Column Splitting: Tools can automatically split each key into a separate column for easy reporting and analysis.
• Automation: JSON data can feed into automated workflows, business intelligence dashboards, or further AI processing without manual parsing of free-form text.
Real-world examples of Summary Prompts
Competitor Analysis
Analyze the {{ session.type }}.
Analyze the conversation between the agent and the client. The client mentioned a competitor company.
Answer the following questions:
- Record which company the client mentioned. Possible competitor company names: “Medlab”, “MedicalCare”, “GlobalLab”, “World Lab”.
- Describe in what context the client mentioned the competitor’s name. Classify the conversation context.
- Determine the sentiment of the mention (positive/negative/neutral).
The answer should be presented in JSON format with the following keys: “Client Mentioned”, “Conversation Context”, “Mention Sentiment”.
The answer should be as short as possible (maximum 50 words) and in {{ language }}.
Advanced Use Case - NPS or Feedback Analysis
Analyze the NPS of the guest's feedback. The following are the characteristics of the product. The answer is in JSON format with the key values below:
- Guest Emotion: Overwhelm, Captivity, Enchantment, Anger, Tasty, Unsavory.
- Type of message: Idea and reason, Rivalry with a competitor, Number of changes, Threat, Information.
- Product and product: which product or product has a seal. Possible options: Casa, Meat, Fish, Confectionery, Pastries, Gastronomy, Dairy, Frozen food, Vegetables, Coffee/tea, Cooking, Alcohol, Drinks/Juice, Groceries.
- The character of the guy: if the comment is to take revenge on the guy, clarify his character. Possible options: Address account, Measurement account, Store account, Subdivision account.
- Category of problem with the product: Rot, Moldy, Third-party item, Added to the taste, Suspect of spoilage, Packing, Taste and freshness, Protermin, Warehouse of the product.
- Assortment: Every day specific product, Request for new product, Incorrectness of surpluses, Small assortment.
- Promotions and prices: High prices, Few promotions, No promotion, Daily promotional item, Price tags, Price tags, Markdown, Insufficient discount.
- Store quality: Smell, Navigation, Parking, Cleanliness, Mosquitoes/rodents, No refrigerators, No cash, Little CSR, Chests, Baskets/carts, Secured passages, Sanitation.
If none of the suggested options fit, propose an answer yourself.