Research brief · 2026-07-13
Retrieval-augmented generation patterns for enterprise knowledge
Four RAG designs AIKm reaches for when a Hong Kong enterprise wants a grounded, cite-able assistant on top of its own documents.
Comparison at a glance
| Pattern | When it fits | Freshness | Cost profile |
|---|---|---|---|
| Classic dense RAG | Text-heavy manuals, policies | Batch, hourly | Low |
| Hybrid (dense + BM25) | Names, codes, product IDs | Batch, hourly | Low–medium |
| Graph-augmented RAG | Relationship-heavy domains (org, product, regulation) | Event-driven | Medium |
| Agentic RAG | Multi-step research, cross-source synthesis | Real-time | Higher |
1. Classic dense RAG
Chunk documents, embed, store, retrieve by similarity, and prompt the model with the top passages. The workhorse pattern: cheap, easy to audit, and good enough for most policy and manual look-ups.
2. Hybrid retrieval
Combine dense embeddings with keyword search (BM25) and rerank. Recovers cases where the user's query contains a product code or person's name that embeddings alone match poorly.
3. Graph-augmented RAG
Extract entities and relationships into a knowledge graph. The agent retrieves both passages and a subgraph, letting answers reason across relationships instead of collating snippets.
4. Agentic RAG
An agent plans a multi-step search, calls retrieval as a tool, and synthesises across sources. Powerful for research questions, but requires the governance controls covered in the agentic AI governance brief.
How to choose
Start with hybrid classic RAG unless the domain is graph-shaped. Add agentic behaviour only where the user need is genuinely multi-step and the governance layer is ready.