Although both engines are designed to identify the user's intent, the new engine incorporates a model that enables it to recognize a wider variety of ways to express the same request.
This approach significantly reduces the need to cover multiple variations of the same question, making the system more flexible and intelligent, allowing you to deliver more accurate and natural interactions to your customers.
How did the classic engine work?The previous engine was based on pattern matching and tokenization (breaking text into words or parts). Intents were configured as text patterns, and the user’s message was split into tokens that were compared against predefined lists.
Based on these matches, the most likely intent was selected. If no suitable match was found, the system would redirect to disambiguation menus or generate fallback responses.
How does the new engine work?The new engine replaces token matching with semantic search based on vectors (embeddings). This means that each intent and its variations are transformed into a numerical representation that captures their meaning.
- Similar meaning → closer together
- Different meaning → farther apart
☝ This is what allows the engine to correctly identify the user’s intent, even when phrased differently.
Once the user query is converted into a vector, the engine searches the database for the closest intent and returns the best match along with a similarity score, indicating the strength of the match.- The minimum score to provide a response is 0.8.
- If the match exceeds this threshold → the associated response is returned.
- If no intent meets the threshold → the system searches within the response base.
- If there is still no result → fallback is triggered: disambiguation or generative AI with RAG (Retrieval-Augmented Generation) to construct a response based on available information.
The new engine not only improves the technology behind your Virtual Assistant, but also transforms how it operates.Here are some of its key benefits:
↑ Higher resolution rate: more conversations are resolved without transferring to a human agent.
↑ Better natural language understanding: interprets multiple ways of expressing the same intent.
↑ Broader query coverage: handles more cases thanks to semantic understanding.
↓ Lower maintenance effort: no need to manually cover every possible variation.
↓ Faster content creation: work with concepts instead of long lists of phrases.