LangThang AI — tour advisory chatbot

What this is, plainly

LangThang AI is a retrieval-grounded tour advisory chatbot for the Vietnamese market, built as a technical project. It is in active development: not launched, not a commercial product, and not a service EthanCorp currently offers. It is on this site because of what it demonstrates, not because it is for sale.

The design question

Conversational advisory systems tend to route everything through the language model because that is the easy architecture. It is also expensive, slower than it needs to be, and unpredictable in exactly the places where predictability matters — dates, budgets, group sizes and availability.

The question this project tests is how much of a real advisory conversation can be handled deterministically, and whether restricting the model to genuine ambiguity produces a better system rather than just a cheaper one.

Constraints

• Vietnamese-language conversation, where off-the-shelf intent tooling is weaker than in English.
• Recommendations must come from real listings — an invented tour is a failure, not a hallucination to be tolerated.
• Hard constraints such as dates, budget and group size are non-negotiable and should never be treated as soft ranking signals.
• The system must not attempt to close a sale; it qualifies and hands off.

Approach

Two-layer NLU. A deterministic rule layer handles clear intents; only ambiguous messages fall through to the model. This keeps latency and cost down, and — more importantly — makes the common path testable.

Structured filtering before vector search. Dates, budget and group size are applied as SQL filters first, so the vector search only ever ranks candidates that already satisfy the hard constraints. Semantic similarity decides which valid option fits best; it never decides whether an option is valid.

Explicit conversation state across four stages — explore, narrow, qualify, hand off. Tracking state explicitly rather than inferring it from history is what makes the closing flow controllable and what makes it testable. 96 backend unit and integration tests cover the rule layer, the filter logic and the state transitions.

Current status

In active development. There is no public launch date, no commercial offering attached to it, and no claim here about production usage — because there is not any yet. When that changes, this page changes.