Research focus
Five connected areas, one programme
Our research runs from how machines find knowledge, to how they structure it, act on it, generate with it, and govern it. Each area feeds the next, and all of it is validated against real enterprise data.
Retrieval-augmented generation (RAG)
Large language models are fluent but forgetful: they know the public internet, not your organisation. RAG grounds model answers in an enterprise's own documents so responses are accurate, current and citable, the foundation of trustworthy enterprise Q&A, search and assistant systems.
Our research examines the full pipeline: document chunking and embedding strategies for mixed English–Chinese corpora, hybrid dense-plus-keyword retrieval, re-ranking, and citation fidelity. We place particular weight on evaluation: frameworks that measure faithfulness and answer completeness, not just fluency.
Current themes: bilingual retrieval quality · long-document strategies · retrieval evaluation benchmarks · hallucination reduction
Enterprise knowledge graphs
Most institutional knowledge is trapped in unstructured content: contracts, reports, emails, wikis. Knowledge graphs extract the entities and relationships inside that content and connect them into a navigable structure: who worked on what, which product depends on which supplier, how decisions trace to evidence.
We study automated graph construction with LLMs, graph-augmented retrieval (GraphRAG) that combines structural and semantic search, and the harder operational problem: keeping a graph current as the organisation, its people and its products change.
Current themes: LLM-based entity and relation extraction · GraphRAG architectures · graph maintenance and versioning · ontology design for business domains
Agentic AI systems
The next layer above retrieval is action: AI agents that plan multi-step work, call tools, query systems and produce finished outputs such as research memos, reports and process runs. The engineering challenge is reliability; the organisational challenge is control.
Our work covers task decomposition, tool-use robustness, multi-agent coordination, and human-in-the-loop checkpoints that let enterprises delegate real workflows without surrendering oversight. We benchmark agent behaviour on realistic enterprise tasks rather than synthetic puzzles.
Current themes: agent reliability measurement · tool and MCP integration patterns · approval and escalation design · agentic workflow economics
Multimodal & generative AI
Enterprise knowledge is not only text. We research generative models across text, image and video for business use, from technical documentation and marketing assets to full AIGC production pipelines for product design and communication.
The questions we pursue are practical: how to keep generated output consistent with brand and factual source material, how to track provenance and comply with emerging disclosure rules, and how the economics of generation at scale change content operations.
Current themes: video generation pipelines · brand-consistent generation · content provenance and watermarking · AIGC workflow design
Responsible AI & governance
None of the above reaches production in a serious enterprise without governance. We develop frameworks for deploying AI in regulated, bilingual business environments: data privacy under Hong Kong's PDPO and cross-border rules, model risk assessment, auditability of AI-assisted decisions, and readiness for international AI regulation affecting companies that operate globally.
This area cuts across every project we run: governance designed in at the start, not retrofitted after an incident.
Current themes: AI risk assessment for SMEs and enterprises · audit trails for agentic systems · regulatory horizon scanning · bilingual policy documentation
Collaborate with usGuide: measuring AI KM ROIAbout the institute
Research briefs
In-depth notes from our engagements
Frequently asked questions
AIKm research focus: FAQ
What does AIKm research?
AIKm researches five connected areas: retrieval-augmented generation (RAG) for enterprise Q&A on bilingual English–Chinese content, enterprise knowledge graphs with LLM-based extraction and GraphRAG, agentic AI systems with human-in-the-loop control, multimodal and generative AI for business, and responsible AI governance under Hong Kong PDPO and international AI regulation.
How does AIKm's research connect to real enterprise deployments?
AIKm's research programme is applied: methods developed in the lab are proven in live enterprise deployments in Hong Kong and the Greater Bay Area. Consulting engagements feed back into research priorities, and every recommendation rests on measured retrieval accuracy, agent reliability and generation quality, not vendor claims.
Which industries benefit from AIKm's research?
AIKm's work applies to finance and insurance, retail and F&B, logistics and freight, property and facilities, and professional services. Bilingual retrieval and PDPO-aware governance make it particularly relevant to Hong Kong enterprises operating across Cantonese, Mandarin and English.
Does AIKm collaborate on research?
Yes. AIKm collaborates with universities, industry bodies (including HKIE) and enterprises on joint research and student placements. Fellow Members Prof Kenneth Lam and Dr W.M. Tsang are both on Stanford University's list of the world's top 2% most-cited scientists.
What is the difference between RAG and a knowledge graph?
RAG retrieves passages of your text and feeds them to a language model so answers cite your documents. A knowledge graph turns the entities and relationships inside those documents into a structured, navigable network. GraphRAG combines both: structural search over the graph plus semantic search over the text, for more complete and traceable answers.