Business users ask questions about internal PDFs, and answers must cite sources. What should you design?
Why C is the answer
A RAG application with citations in the output
RAG retrieves the source passages, so answers can stay grounded and cite them.
In detail
- The answers must come from the PDFs, and they must cite them.
- RAG retrieves the passages first, so the answer can stay grounded and name its source.
- Return the source with each passage, so citations come from real retrieval.
Use case
An illustrative scenario, not a real incident.
An insurer's staff ask about thousands of policy PDFs. A RAG app retrieves the passages, answers from them and links the document and page, so staff can check the source.
See it in different ways
flowchart TD
A{"What kind of knowledge?"} -->|"Policies and manuals"| B["Vector Search"]
A -->|"Facts for one order or customer"| C["Lookup by key in a table"]
A -->|"Live values"| D["A tool call"]mindmap
root((RAG app))
Data
Delta tables
Chunks
Retrieval
Vector index
Reranking
Chain
MLflow
Agent Framework
Serving
Model Serving
Governance
Unity CatalogsequenceDiagram participant P0 as User participant P1 as RAG app participant P2 as Vector index participant P3 as LLM P0->>P1: Question P1->>P2: Find passages P2-->>P1: Passages with sources P1->>P3: Answer from these P3-->>P0: Answer with citations
Key point: Citations need retrieved passages to point at.
Why the other options are wrong
- A. A nightly BI dashboard built from the extracted PDF text
- A dashboard shows fixed views. It does not answer new questions.
- B. A plain classification model that labels each question
- A classifier labels the question and never produces an answer.
- D. A fine-tuned model with no retrieval step
- A fine-tuned model has no passages to cite, and it goes stale when documents change.