Singapore · RAG accuracy · Evidence you can audit
QCRCE measures and improves retrieval accuracy for RAG and internal search at Singapore and Southeast Asia tech teams. We test on your own documents and real questions, report recall@k and citation accuracy, and agree pass/fail acceptance tests with you before we build.
A RAG system that answers confidently from the wrong passage erodes trust fast. Most teams cannot say what share of answers are grounded in the right source.
Reviewers, risk teams and — for financial institutions — the MAS Guidelines on AI Risk Management (effective 7 October 2027) expect AI to be evaluated against measurable thresholds and documented.
Staff and customers in the region ask in Chinese, Malay and Bahasa. Retrieval that works in English can fail silently in other languages.
Build a gold set from your real questions (typically 50) with the source passage for each answer. Baseline recall@k and citation accuracy on your current system.
Before any build work, write down the acceptance tests together: which metrics, which thresholds, which languages. No moving goalposts.
Improve chunking, embeddings, hybrid search or re-ranking, re-run the same tests, and deliver a go / no-go report with the evidence behind it.
Fixed scope, fixed price. Prices in SGD; invoiced via our Singapore business account.
Entry offer
SGD 3,800 about 1 week
Full prepayment before work starts.
Package A · Pilot
SGD 8,000 2–3 weeks
50% deposit before kickoff, 50% on delivery.
Package B
On request 2–4 weeks
Paid trial. 50% deposit before credentials are issued.
Acceptance tests are agreed per project, before building. As an example only, a pilot might target recall@10 ≥ 0.85, citation accuracy ≥ 0.80 and p95 latency ≤ 200 ms.
These are example targets, not results from a customer project.
Our public demo runs this method on MAS AI risk documents with a 32-question gold set and publishes every measured number, including the misses.
For financial institutions and their technology vendors, the MAS Guidelines on AI Risk Management (finalised 7 October 2026, effective 7 October 2027) call for evaluation against measurable thresholds, documentation that lets an independent reviewer reproduce the work, and ongoing monitoring.
Our deliverables are designed to support your evidence pack: the gold set, test results, failure analysis and run instructions can be filed with your AI inventory and review records.
QCRCE provides software and technical evaluation only. We do not provide legal, regulatory or compliance advice or certification; your compliance and legal teams decide what meets your obligations.
Some Singapore companies may be eligible for Enterprise Singapore's EDGE Grant (which replaced EDG and PSG from 30 September 2026). Eligibility and approval are decided by Enterprise Singapore, and the grant generally requires that no work or payment happens before you apply.
We are glad to provide a quotation and scope for your own application, but we cannot promise any subsidy. If you plan to apply, tell us first and we will not start or invoice until you have.
Your data is used only to deliver the contracted package and is never used to train public models. Third-party APIs are disclosed in the SOW, with SG/SEA regions preferred. Corpora are deleted or returned within 30 days of project close.
QCRCE is a small Singapore software and technical consulting company focused on RAG accuracy, retrieval evaluation and vector retrieval. It is founded and run by Mark (Ma Weijia), who works directly on every engagement.
We are software and consulting only. We do not provide financial or regulated advisory services.
| Legal name | QCRCE Pte. Ltd. |
|---|---|
| UEN | 202630091K |
| Address | 68 Circular Road #02-01 Singapore 049422 |
| Founder | Mark (Ma Weijia) |
Tell us what your RAG or search system needs to answer, and where the knowledge lives. We will reply with whether a package fits.