Knowledge
Knowledge HubConcepts and strategic researchGlossaryDefinitions across six domainsFrameworksStructured strategy methodsTopicsConnected subject areasLibrary
PlaybooksPractical strategy guidesBooksBooks and long-form researchOperator VaultMember resources and toolkits
Explore Discover, the Knowledge Hub, playbooks, and the Operator Vault.
Glossary
Retrieval-augmented generation
Retrieval-augmented generation (RAG) is a pattern that combines a generative model with a retrieval step: the system fetches relevant external passages (or records) and conditions generation on that non-parametric context, rather than relying only on the model’s parametric memory. The influential formulation appears in Lewis et al. (2020).
Why it matters
Generative models can produce fluent answers that are incomplete, outdated, or unsupported when required knowledge is missing from parameters or the prompt. RAG addresses many of those cases by supplying retrieved enterprise or web content at generation time — improving updateability and domain specificity when retrieval quality is high. It does not automatically make every answer correct.
Separate four concerns in evaluation: (1) whether the right sources were retrieved, (2) how generation used that context, (3) whether citations match the passages shown, and (4) whether the claim is factually verified. Strong citation formatting can still accompany weak retrieval or unsupported conclusions.
Example (enterprise): A revenue-operations assistant retrieves the current discount policy and CRM snippets for an account, then drafts a deal desk note. If the index is stale or ACL filters omit the latest policy, the draft can be wrong even though the LLM is generating fluent text.
Mechanisms
- Retrieve — select top passages via semantic search, hybrid search, or structured lookup.
- Augment — insert retrieved context into the prompt or model conditioning.
- Generate — produce an answer, draft, or action proposal that may use that context (and citations/guards when implemented).
Boundaries
- RAG ≠ semantic search alone (search may not generate).
- RAG ≠ fine-tuning; both can coexist but solve different update and behavior problems.
- RAG ≠ a guarantee of factuality; retrieval errors, conflicting sources, prompt misuse, and unverified claims still produce bad outputs.
Related topics
None linked yet.
Related concepts & frameworks
Related Discover collections
Further research
No related resources yet.