Can JEV / TEV replace Embedding for intent recognition? | Jev / TEV 能不能干掉 Embedding 做意图识别？

Jev and TEV may not fully replace traditional embedding-based intent recognition, but they can replace many scenarios. Both approaches have defined roles, not simply one replacing the other. To understand the essence of each, let's break down the two systems:

1. Traditional embedding approach uses vectors to recognize intent. The process involves: sentence → vectorization → vector similarity search, matching a predefined intent library. The principle is semantic similarity matching based on vector proximity to determine the corresponding intent. It's suitable for fixed, pre-built intent libraries with large numbers (thousands to tens of thousands of intents).

The drawback is that it only considers semantic similarity and lacks understanding of logical conditions and constraints. For example, the vector for "I want a refund but the order has been over 180 days" might match the "refund" intent, but the business rules don't allow refunds. It's easy to misinterpret sentences with negations: "I don't want to refund" still gets matched to the refund vector.

Adding new intents requires reindexing and adding samples, which is difficult. Scaling up to parallel judgment of multiple intents is challenging for embedding.

2. Jev/TEV, the System-One decision model, handles intent recognition. Choice/Noul does native intent classification with calibrated probability output. It understands the description, conditions, negations, and boundary definitions. The route_team structure allows direct input of business boundary conditions in the criteria section, which embedding cannot achieve.

Advantages include zero-shot/low-shot learning, allowing direct addition of new intents without pre-built vector libraries. New intents are simply added as JSON configurations without reindexing. It naturally supports negations, conditions, and complex business rules. It outputs parallel results: intent, urgency, and risk, all with calibrated probabilities for setting thresholds and automatically passing to System2 if confidence is low.

However, it has a limitation: a hard upper limit of 255 options. If your intent library has 2000+ fine-grained intents, TEV/Jev cannot handle it. The cost increases with large numbers of long-tail intents: several hundred intents are manageable, but a few thousand require maintaining numerous criteria texts, raising inference costs.

It has limited generalization for completely new, unseen intents, unlike the recall ability of embedding systems.

TEV/Jev can replace embedding in the following scenarios:

- Intent count is between tens and a few hundred

- Business boundaries between intents are complex with many conditions, exceptions, and negations

- Frequent additions or modifications to intents, avoiding constant vector library maintenance and re-indexing

- Need for simultaneous parallel judgment of urgency, risk, and truthfulness

- Scenarios like customer service ticket triage, agent tool selection, or risk screening in fraud detection

Embedding should remain for scenarios with:

- Large intent counts: several hundred to over a thousand fine-grained intents, such as enterprise FAQ systems with 3000+ answers. Embedding excels at recalling historically similar user queries, which is optimal for retrieval tasks rather than decision classification. Embedding is indispensable for knowledge base retrieval, similar query recall, and RAG systems.

- Large-scale scenarios with long-tail intents requiring coarse screening recall followed by decision model refinement.