Open-source live-chat, email support, omni-channel desk. An alternative to Intercom, Zendesk, Salesforce Service Cloud etc. 🔥💬
Ruby actioncablechat-widgetconversationcustomer-supportdashboarddesigndockerdocker-imageherokuintercomjavascriptlivechatopensourcerailsrubysassvuejswhatsappzendesk
class Captain::ConversationClassifierService
MODEL = 'jev-latest'.freeze
MAX_LABEL_SUGGESTIONS = 3
LABEL_THRESHOLD = 0.5
PRIORITY_CRITERIA = {
'urgent' => 'The customer is blocked right now with serious impact: an outage, failing payments, a security issue, data loss, ' \
'or a hard deadline within hours.',
'high' => 'A significant problem affecting the customer\'s work or money that needs attention soon, but is not an emergency.',
'medium' => 'A normal support request or question with no sign of time pressure.',
'low' => 'Feedback, general information, or a casual message that can wait.'
}.freeze
LABEL_CRITERIA = {
'true' => 'The messages are clearly about the topic this label names or describes.',
'false' => 'The messages are about something else, or only mention the topic in passing.'
}.freeze
class Error < StandardError; end
pattr_initialize [:conversation!]
def labels
candidates = conversation.account.labels.where.not(title: conversation.label_list).to_a
return [] if candidates.empty?
answers = ask(candidates.each_with_index.to_h { |label, index| [index.to_s, label_question(label)] })
candidates.each_with_index
.map { |label, index| { title: label.title, probability: answers[index.to_s]['noul'] } }
.select { |suggestion| suggestion[:probability] >= LABEL_THRESHOLD }
.max_by(MAX_LABEL_SUGGESTIONS) { |suggestion| suggestion[:probability] }
end
def priority
answer = ask(
priority: {
type: 'choice',
instructions: 'How urgently does the support team need to respond to the customer in `conversation.messages`?',
criteria: PRIORITY_CRITERIA
}
)['priority']
{ priority: answer['choice'], confidence: answer['confidence'] }
end
private
def label_question(label)
{
type: 'noul',
instructions: {
label: { name: label.title, description: label.description }.compact_blank,
question: 'Does `label` describe what the customer is asking about or reporting in `conversation.messages`?'
},
criteria: LABEL_CRITERIA
}
end
def ask(questions)
state = { conversation: { messages: Captain::ConversationTranscript.new(conversation: conversation).messages } }
body = Captain::JevClient.request_body(model: MODEL, state: state, questions: questions)
Captain::JevClient.new(account_id: conversation.account_id, conversation_id: conversation.display_id,
feature: 'captain_classifier').call(body: body)['answers']
rescue Captain::JevClient::HTTPError => e
raise Error, e.message
end
end