Dialogflow Knowledge Connectors

A recurring challenge in chatbot development: enabling systems to answer numerous questions on specific subjects by leveraging existing knowledge bases housed in FAQ pages, PDFs, or unstructured documents.

The Problem

These repositories often contain thousands of potential answers, yet the primary challenge remains extracting relevant information from unstructured sources to provide accurate responses.

The Traditional Approach

Conventional chatbot frameworks including Google Dialogflow, IBM Watson, Microsoft Bot, and Rasa operate similarly. User queries are matched to intents with entity extraction, then either static responses are provided or application layers process requests. This approach becomes problematic when supporting large question scopes or frequently updated information, leading to intent classification model degradation, substantial effort maintaining training data accuracy, and ongoing intent creation and management overhead.

Knowledge Connectors

Released in 2019 as a beta feature, Knowledge Connectors supplement traditional intent matching. When queries don't match intents, the system consults configured knowledge bases containing documents (text/csv, text/html, application/pdf, plain text) or web URLs.

Enable beta features in agent settings. Configuration occurs through the web console or client libraries (Java, Node.js, Python). Create knowledge bases by navigating to the knowledge tab and providing a name plus document. Multiple response types, including rich responses, are supported.

Testing: FAQ Knowledge Base

Testing used the UCAS Frequently Asked Questions webpage. The system successfully processed the URL, creating extractable question/answer pairs. Initial queries performed exceptionally: "how do I apply" returned matchConfidenceLevel: HIGH with 0.973 confidence. However, less obvious queries produced incorrect matches. "How do I submit my application" matched the wrong intent despite high confidence (0.962).

A valuable feature allows assigning specific extracted FAQs to intents via the "convert to intents" button, automatically creating new intents while disabling corresponding question/answer pairs.

Testing: Unstructured Document FAQ

Testing employed a Priorix medication leaflet PDF. Despite high confidence scores, extracted answers proved inaccurate. Results suggested keyword-based matching without contextual consideration. No mechanisms exist to evaluate or train responses, limiting control compared to traditional intents.

Do Knowledge Connectors Work?

For existing FAQ pages or similarly structured documents, Knowledge Connectors demonstrate solid performance. However, monitoring through history logs remains essential for identifying mismatches. For unstructured documents requiring answer extraction, they don't represent a comprehensive solution. While Knowledge Connectors are experimental, improvements may emerge as technology advances.