Build Retrieval-Augmented Generation (RAG) pipelines with modules for document ingestion, embedding, retrieval, and grounded LLM responses.
The foundational pattern for answering questions from a document corpus: chunk and embed the documents, retrieve the passages relevant to a question, and generate an answer constrained to what was retrieved. Everything else in this group is either a variant of this or built on top of it.
A body of documents too large to read, questions that cannot be anticipated in advance, and answers that must be traceable to source.
Questions requiring computation over structured data — RAG will retrieve passages about the numbers rather than calculating with them. Also poor where the answer requires reading the whole corpus rather than a few passages.
Retrieval must respect the asking user's permissions. A pipeline that embeds everything into one index and retrieves without access filtering will happily answer from documents the user could not otherwise open.
Retrieval failure presented as an answer. When nothing relevant is retrieved, the model generates from parametric knowledge instead of saying it does not know, and the output looks identical to a grounded one.
An auditor asking what supports a specific balance, where the system retrieves the reconciliation, its evidence documents and the correcting entry — and returns nothing rather than improvising when the account is outside the requested period.