Talent policy is a moving target — subsidies, housing, settlement and qualification recognition come from different departments, tiers and effective dates. The group built a policy RAG chatbot on MOI whose answers carry their source and effective date.
Shenzhen Talent Group is the municipal government's comprehensive talent-services platform, wholly state-owned under Shenzhen SASAC and a central instrument of the city's talent-strong strategy and the Greater Bay Area talent hub. It covers the full HR services value chain, from executive search to public talent services and HR outsourcing.
Shenzhen's talent policy is a moving target: subsidy schemes, housing support, settlement and qualification recognition, funding programs-issued by different departments, revised frequently, each with its own eligibility conditions, required materials and application windows. Hundreds of thousands of users-graduates, job seekers, returning overseas talent and enterprise HR teams-ask the same questions through hotlines, service windows and WeChat channels. Frontline staff answered by hand from documents scattered across sources, so answers varied from person to person, policy revisions took time to reach the people answering, and demand spiked exactly when policies were released. In a public-service setting the cost of a wrong or outdated answer is real-an applicant misses a deadline or files against a policy that no longer applies.
Shenzhen Talent Group deployed a talent-policy RAG chatbot on OmniFabric. Municipal and district policies, implementation rules, application guides and historical Q&A are parsed and chunked at clause level, embedded, and indexed in the same engine that holds the structured service data-so vector embeddings and the full-text index live together instead of in separate systems. Retrieval is hybrid: keyword matching pins exact policy names and article numbers while semantic search handles the way citizens actually phrase questions. Answers are generated with the source document and its effective date attached, superseded versions are excluded by policy versioning, and low-confidence questions route to a human agent.
Hundreds of thousands of users get talent-policy answers on demand across web, mini-program and hotline channels, without adding frontline headcount every time a policy is released.
Every answer is generated from the current policy corpus with the source document and effective date attached, so guidance no longer varies with whichever staff member picks up the question.
Revised policies enter the corpus and take effect in answers at once while superseded versions are excluded, closing the lag between issuance and what the frontline knows.