Enterprise RAG · Hong Kong
RAG Implementation in Hong Kong
Most of what is written about Hong Kong RAG is written by people who have not deployed one here. We have shipped three: an on-premise retrieval system over a Big 4 firm's compliance library, a private platform for a family office where nothing leaves the room, and a multimodal RAG application over text, tables and images.
Bilingual English and Traditional Chinese retrieval, self-hosted so the corpus stays inside your perimeter, and every answer arrives with the passage that supports it. In audit and investment work, an answer you cannot cite is not an answer — it is a liability.
Expert-Vetted top 1% · 100% JSS · production AI at a 5,000-employee US enterprise + a Big 4 firm
The Proof
Three RAG Systems We Shipped
Not three blog posts about RAG. Three retrieval systems running on client infrastructure, with the constraints that shaped them.
Enterprise RAG Chatbot for a Big 4 Accounting Firm
Big 4 client · fully on-premiseJunior auditors at a Big 4 firm were reading thousands of pages of compliance whitepapers and audit guidelines to answer a single question. Every senior became a search engine, and the same question got asked and re-answered across teams that never saw each other’s work.
A multimodal retrieval system over the full compliance library, running entirely on the firm’s own infrastructure. OCR and table parsing mean the semi-structured exhibits — the rate tables, the disclosure matrices, the scanned annexes — are searchable rather than silently skipped. Retrieval is tuned for citation rather than summary, so the system returns the passage that supports the answer alongside the answer itself.
In audit work, an answer you cannot cite is not an answer — it is a liability. That single constraint shaped the whole build: the ranking, the chunking strategy, the refusal behaviour when the library genuinely does not cover the question. A confident guess is worse than no answer at all when the output ends up in a working paper.
Super Chain: Private On-Prem AI for a Family Office
Nothing leaves the roomA family office holds the ledger. Fund statements, trust deeds, counterparty terms and tax filings are exactly the documents that cannot be posted to somebody else’s API. That is why most offices of this kind had no AI at all — not because the use case was weak, but because every available product required handing the crown jewels to a third party.
Super Chain: a fine-tuned LLM with retrieval over financial research, macro data and portfolio holdings, running on the office’s own infrastructure. It is fine-tuned on the domain rather than prompted at it, so it speaks the register the analysts actually use, and every answer is composed with its sources cited so a principal can check the working.
Nothing leaves the room. The privacy constraint is not a feature bolted on at the end — it decided the model, the deployment and the retrieval architecture from the first design session. That is the difference between an AI system a family office can adopt and one their counsel will never sign off on.
Multimodal RAG Application
246 hours · sustained engagementA retrieval application over multimodal data, where the source material is part text, part table and part image rather than the clean prose most RAG tutorials assume.
Each format parsed on its own terms before anything reaches the vector store, rather than flattening everything to text and hoping the embedding model sorts it out. 246 hours of sustained engineering across the ingestion, retrieval and evaluation layers.
That mix of formats is where naive retrieval quietly fails: the chunker splits a table down the middle, the image carries the number nobody indexed, and the answer comes back confident and wrong. Nothing errors, nothing alerts — you simply get wrong answers delivered with total fluency, which is the worst failure mode a retrieval system has.
Full detail on all eight engagements: production AI case studies.
Why RAG Returns Wrong Answers
Four Silent Failures in Enterprise RAG
Practitioner pain has moved from generation to retrieval. Every one of these fails quietly — nothing errors, nothing alerts, the answer just comes back wrong and fluent.
The table gets split down the middle
Hong Kong enterprise documents are rate tables, disclosure matrices and scanned annexes — not the clean prose RAG tutorials assume. A naive chunker cuts a table in half and the retriever confidently returns the wrong half. Nothing errors. You get wrong answers delivered with total fluency.
The query is in English, the document is in Chinese
A board pack in English cites a filing in Traditional Chinese; a policy exists in one language with no authoritative translation. Single-language embeddings simply never retrieve the other half of your estate. Cross-lingual retrieval has to be an architecture decision, not a translation step bolted on at query time.
The answer has no source attached
In audit, compliance and investment work, an answer you cannot cite is not an answer — it is a liability. Retrieval tuned for summary produces fluent paragraphs nobody can defend in a working paper. Retrieval tuned for citation returns the supporting passage alongside the answer.
Nobody built the evaluation set
Teams demo, ship, then discover edge cases through production failures. A labelled evaluation set built before the first deployment and re-run on every release is what turns a demo into a system. It is the cheapest thing to build and the most expensive thing to skip.
Regulated Hong Kong
What Changes When the Corpus Is Confidential
For a bank, an insurer, a family office or a professional-services firm, four constraints decide the architecture before anyone picks a model.
Data residency
Client data under the PDPO and HKMA cross-border expectations. Self-hosted retrieval and a self-hostable embedding model keep the corpus inside your perimeter — the deciding constraint for both the Big 4 and family-office builds.
Explainability
The HKMA’s GenAI guidance expects an institution to explain an AI-assisted output. Citation-first retrieval is the practical answer: the source paragraph ships with the answer, so a reviewer can check the working.
Human oversight
High-risk outputs need a review gate. The system should refuse rather than guess when the library genuinely does not cover the question — refusal behaviour is a design decision, not a fallback.
Audit logging
Financial records carry long retention expectations. Log the query, the retrieved evidence and the generated answer, so a disputed output six months later is reconstructable.
General guidance on how these constraints shape a retrieval build — not legal or regulatory advice. Confirm the applicable requirements with your own risk and compliance function.
The Landscape
Who Does RAG in Hong Kong, and When to Pick Them
An honest map, including where somebody else is the better call. We are not the right answer for a clean, single-language, customer-service FAQ bot.
| Who | What they do | Pick them when |
|---|---|---|
| HKPC (Hong Kong Productivity Council) | Institutional programmes — Agentic AI Foundry, the PistaChio knowledge portal, SME-facing AI showcases. | Funding, training and a first look at the technology. Not a team that embeds in your codebase. |
| ThinkCol | A local GenAI consultancy with a RAG chatbot platform, sold across retail, property, finance and government. | You want a platform and a project team. You accept the platform’s retrieval architecture as given. |
| GPTBots / Aurora Mobile, Wavenex, MasterConcept | Upload-your-documents chatbot products with RAG underneath, largely SME and customer-service shaped. | The corpus is clean, English-ish, and the answer does not have to survive an audit. |
| Offshore agencies publishing HK guides | Detailed English-language content on HK and Singapore enterprise RAG, written from outside the market. | You are researching. Check whether the named case studies are theirs before you shortlist them. |
| Super Cat Technology | Embedded top-1% engineers who have already shipped on-premise RAG for a Big 4 firm and a family office. T&M, your infrastructure, no platform licence. | The corpus is bilingual, semi-structured and confidential, and every answer needs a citation behind it. |
FAQ
Hong Kong RAG, Answered
Who builds RAG systems in Hong Kong?
The Hong Kong market splits three ways. Institutional programmes — HKPC’s Agentic AI Foundry and PistaChio, and academic work such as HKU’s RAG-Anything — build capability and awareness. Local consultancies and product vendors, including ThinkCol, GPTBots via Aurora Mobile, Wavenex and MasterConcept, sell RAG chatbot platforms where you upload documents and get an assistant. And a small number of firms embed senior engineers who build the retrieval system on your own infrastructure. Super Cat Technology is in the third group: we have shipped on-premise enterprise RAG over a Big 4 accounting firm’s compliance library and a private retrieval system for a family office, plus a multimodal RAG application over mixed text, tables and images. We work time-and-materials from Hong Kong and London, with no platform licence.
Can a RAG system handle Traditional Chinese and English documents together?
Yes, but only if it is designed for it. Hong Kong enterprise corpora are genuinely bilingual — board minutes in English referencing filings in Traditional Chinese, policies that exist authoritatively in one language only. A single-language embedding model will never retrieve the other half of the estate, and translating at query time loses the terminology that matters in financial and legal text. The working approach is a multilingual embedding model that supports both dense and sparse retrieval, language-aware chunking, and a hybrid layer that combines vector search with keyword matching so exact identifiers — a policy number, an ordinance reference — still hit exactly.
Can our data stay in Hong Kong, on our own servers?
That is the normal case for the work we do, not the exception. Both the Big 4 compliance system and the family-office platform run entirely on the client’s own infrastructure, because the documents involved — audit guidelines, fund statements, trust deeds, counterparty terms — are exactly the documents that cannot be posted to somebody else’s API. Self-hostable embedding models and vector stores make a fully on-premise retrieval stack practical, and for PDPO and HKMA-supervised institutions that constraint should shape the architecture from the first design session rather than be retrofitted.
Why do RAG systems return wrong answers even when the document is in the index?
Almost always retrieval, not generation. Four failure modes account for most of it: chunking that splits a table or a scanned exhibit so the retrieved fragment is missing the number; a query in one language against documents in another; retrieval tuned for summary so the answer arrives without the passage that supports it; and no evaluation set, so nobody knows the failure rate until a user finds it. The fix is unglamorous — parse each format on its own terms before anything reaches the vector store, filter metadata so superseded document versions cannot be retrieved, combine vector and keyword retrieval, and build a labelled evaluation set before the first production deployment.
What does a Hong Kong RAG implementation involve, and how long does it take?
Scoping and a document-estate review first — what formats, how many languages, which versions are authoritative, who is allowed to see what. Then an ingestion and retrieval build tuned to that estate, an evaluation set built before anyone else sees the system, and instrumentation so failures are findable. Timelines depend on the corpus rather than the model: a clean single-language library moves fast, while scanned bilingual filings with permission boundaries take longer in ingestion than in anything else. We run these as embedded time-and-materials engagements, so you scale the team with the work instead of buying a fixed scope.
Should we buy a RAG platform or build our own?
Buy when your corpus is clean, single-language, non-confidential and the answers do not have to survive an audit — a platform will get you there faster and cheaper. Build when the constraints decide the architecture: data that cannot leave your infrastructure, bilingual retrieval, semi-structured exhibits, or a citation requirement that a general-purpose chatbot cannot meet. We are vendor-neutral on this and will say so when a platform is the right answer; we make money on engineering time, not on licences.
Bring us the corpus everyone else said was too messy.
Scanned bilingual filings, versioned policies, tables nobody can index. That is the work we have already done twice on client infrastructure.