Why does a model need external knowledge?
Model parameters are not a continuously updated enterprise knowledge base. Manuals, contracts, and internal rules change, and different users may be allowed to see different documents. RAG separates storage from response generation: it retrieves currently available material for the question, then asks the model to synthesize an answer. Knowledge can therefore be updated without retraining the model for each change.
What happens between indexing and answering?
An offline pipeline collects documents, parses them, splits them by meaning or structure, and builds keyword, vector, or hybrid indexes. Online, a question triggers retrieval of candidate passages. The system may rerank them before passing selected passages and provenance to the model. After generation, the application can check citations, record what was retrieved, and acknowledge insufficient evidence. Enterprise systems also need access control and document versioning at indexing and retrieval time.
For an expense policy question, RAG can find the current clause, explain its scope, and link to the original. Regional rules require the employee's location. Conflicting documents require version and owner checks rather than an arbitrary model choice. Multimodal retrieval and analytics illustrates a retrieval workflow feeding later analysis.
Where does RAG help, and where can it fail?
RAG suits large, frequently updated collections where provenance matters. Its quality depends on parsing, chunking, indexing, recall, ranking, and generation. Missing the right document, retrieving a stale version, or disclosing material outside a user's permissions can undermine the answer even when it has a citation. Structured metric calculations often need a semantic layer and query engine rather than inference from prose alone.
FAQ
Does RAG require a vector database?
No. Keyword search, database retrieval, and hybrid methods can also supply candidates.
How does RAG differ from fine tuning?
RAG reads external material at answer time; fine tuning changes model behavior through training. They may be combined.
How should RAG be measured?
Evaluate evidence recall, citation accuracy, answer completeness, unanswerable questions, and permission boundaries separately.