Reviews, surveys and tickets

The conversation format was built for chats, but it fits any customer text. The rule: anything written by the customer or someone outside your company is incoming: true. Anything your company wrote is incoming: false.

Reviews

One review is one row, and one conversation.

Column	Value
sessionId	The review id
userId	The reviewer id, or the review id if there isn’t one
incoming	true
text	The review body
extra columns	Star rating, location, source, review date

If you also have your company’s replies, add each reply as a second row with the same sessionId and userId, with incoming: false and the reply’s own time.

Drop reviews with no text rather than filling them with placeholder text such as [No review text]. A placeholder looks like a real review to the model. Keep the star rating in an extra column: Dimensions are read from the text only, so the model never sees the stars.

Surveys

Each respondent is one conversation, with the response id as sessionId. Each question your company asked is an outgoing message, and each answer is an incoming one.

Row	incoming	text
1	false	Why did you give that score?
2	true	The checkout kept failing on my phone.
3	false	What one thing should we improve?
4	true	Fix the app.

Keep the scores (NPS, CSAT, ratings) as extra columns. The converter turns a wide survey export (one row per respondent, one column per question) into this shape, and can take the question text from the row of question wording many survey tools put under the column names. Use the wording respondents saw, not column codes such as Q3_why. For more control, see Reshape a wide survey.

Tip

Because the model reads only the text, it can’t see which question a short answer like “Fix the app” responds to unless the question is in the session too. That’s why the questions go in as outgoing rows.

Tickets and email threads

One ticket is one conversation. Each message in the thread is a row, incoming: true for the customer and false for your agents. Use the ticket id as sessionId and conversationId, so the thread stays together when different agents reply. Use the customer’s id as userId, on the agents’ replies too (the userId rule). Strip HTML and quoted replies first: see Clean ticket and email text.

Call transcripts

One call is one conversation, with the call id as sessionId and one row per speaker turn. Turns need timestamps at least one second apart, in the order they were spoken.

Several channels in one place

You can upload chat, email, reviews and surveys as separate sources, or together under the Omni Channel type. Either way, add a channel column so the agent can compare them.

Templates

Start from a template: reviews, survey answers or chats and tickets.