iachat/enterprise/app/services/captain/llm/assistant_chat_service.rb
Pranav 6096932f76
feat: Add a review step for FAQs generated from conversations before using it (#10693)
This PR introduces a review step for generated FAQs, allowing a human to
validate and approve them before use in customer interactions. While
hallucinations are minimal, this step ensures accurate and reliable FAQs
for Captain to use during LLM calls when responding to customers.

- Added a status field for the FAQ
- Allow the filter on the UI.
<img width="1072" alt="Screenshot 2025-01-15 at 6 39 26 PM"
src="https://github.com/user-attachments/assets/81dfc038-31e9-40e6-8a09-586ebc4e8384"
/>
2025-01-16 09:54:34 +05:30

103 lines
2.5 KiB
Ruby

require 'openai'
class Captain::Llm::AssistantChatService < Captain::Llm::BaseOpenAiService
def initialize(assistant: nil)
super()
@assistant = assistant
@messages = [system_message]
@response = ''
end
def generate_response(input, previous_messages = [], role = 'user')
@messages += previous_messages
@messages << { role: role, content: input } if input.present?
request_chat_completion
end
private
def system_message
{
role: 'system',
content: Captain::Llm::SystemPromptsService.assistant_response_generator(@assistant.config['product_name'])
}
end
def search_documentation_tool
{
type: 'function',
function: {
name: 'search_documentation',
description: "Use this function to get documentation on functionalities you don't know about.",
parameters: {
type: 'object',
properties: {
search_query: {
type: 'string',
description: 'The search query to look up in the documentation.'
}
},
required: ['search_query']
}
}
}
end
def request_chat_completion
response = @client.chat(
parameters: {
model: DEFAULT_MODEL,
messages: @messages,
tools: [search_documentation_tool],
response_format: { type: 'json_object' }
}
)
handle_response(response)
@response
end
def handle_response(response)
message = response.dig('choices', 0, 'message')
if message['tool_calls']
process_tool_calls(message['tool_calls'])
else
@response = JSON.parse(message['content'].strip)
end
end
def process_tool_calls(tool_calls)
process_tool_call(tool_calls.first)
end
def process_tool_call(tool_call)
return unless tool_call['function']['name'] == 'search_documentation'
query = JSON.parse(tool_call['function']['arguments'])['search_query']
sections = fetch_documentation(query)
append_tool_response(sections)
request_chat_completion
end
def fetch_documentation(query)
@assistant
.responses
.approved
.search(query)
.map { |response| format_response(response) }.join
end
def format_response(response)
"\n\nQuestion: #{response[:question]}\nAnswer: #{response[:answer]}"
end
def append_tool_response(sections)
@messages << {
role: 'assistant',
content: "Found the following FAQs in the documentation:\n #{sections}"
}
end
end