A customer-support assistant needs current, specific answers, but its training data may be outdated or lack internal documentation altogether. A knowledge base supplies documents the assistant can retrieve as evidence instead of relying only on what the model remembers. section 1.2
In a RAG system, the knowledge base sits between an offline indexing pipeline and an online query pipeline. Indexing breaks documents such as manuals, policies, and FAQs into manageable pieces, converts them into searchable representations, and stores them. At query time, retrieval fetches candidate pieces; ranking prioritizes useful matches before the language model generates an answer grounded in that evidence. section 1.2
An acronym collision
A fintech knowledge base may span product wikis, runbooks, compliance guides, and API specifications. “CMA” can name different things across those documents. Retrieving the wrong CMA document is not a lack-of-content problem: the right information may exist, but retrieval has missed the best match. In one baseline using query reformulation, single-pass vector retrieval, and answer summarization, Hit Rate @5 was 54.12% on 85 questions, with average query latency of 0.79 seconds. Procedural and multi-document questions were also difficult. section 15.7.1
Keeping knowledge in a document repository also makes it possible to update information without retraining the whole model and to inspect which document supported an answer. Neither benefit guarantees a correct response: the relevant document still has to be retrieved, ranked well, and used accurately during generation. section 3.4 section 1.2














