Cantonese and code-switching
A sentence in Hong Kong moves between Cantonese, English and written Chinese mid-clause. The model is evaluated on that, not on single-language text.
Language models
Language models adapted for Hong Kong.
WellBridge LLM is the language-model programme: large language models adapted for Cantonese, Traditional Chinese, English, Hong Kong terminology, code-switching, grounded organisational knowledge and secure deployment.
Research areas
What the language models are built to handle.
Traditional Chinese and English
Both scripts and both languages, using the terminology a Hong Kong professional actually writes, rather than a translated approximation of it.
Grounded organisational knowledge
Answers come from documents the client has approved, and the source is shown with the answer so a reader can check it rather than trust it.
Client separation
One client's knowledge base is never reachable from another's, and recall is per task or opt-in per project. There is no unrestricted cross-client memory.
Secure deployment
Versioned endpoints, structured-output enforcement, retry limits and a prompt registry, inside the client's own tenancy where that is required.
How a request is handled
Ask, retrieve, answer, and stop when the answer is not supported.
AskThe request arrives in the language it was written in.
RetrieveOnly approved sources are searched, inside the client's own boundary.
Answer with sourcesThe reply carries the documents it came from, so it can be checked.
EscalateWhere the sources do not support an answer, the model says so and a person picks it up.
Evaluation
What is measured, and how.
- 01 Local terminology and code-switching accuracy.
- 02 Groundedness: whether a claim is actually carried by the source it cites.
- 03 Citation correctness, checked against the source text.
- 04 Refusal accuracy in both directions — it should decline what it cannot support, and not decline what it can.
- 05 Latency and cost per completed task.
- 06 Completion of the reader's actual task, on consented Hong Kong evaluation sets.
What this is not
It is not a general-purpose chatbot, and this site does not present it as one.
It does not answer from the open internet unless a deployment is configured to, and never without showing what it used.
Public third-party benchmarks are reported separately from the business-task figures and are never averaged together.
Next step
Start with a conversation.
A consultation is a scoped conversation about what you need, not a sales call. You will leave it knowing whether this is the right practice for the work.