Infrastructure Decisions · judgement tool

RAG or a long context window?

Corpus size vs the context window, update frequency, citations and per-call cost decide it. Get a verdict with each factor tagged official fact / calculation / StackSays rule / estimate.

A judgement tool, not a tool ranking · thresholds are StackSays rules, not industry standards · official facts verified 2026-07-17 · share the URL to share your answers

How to read this: every factor below is tagged by where it comes from — official fact official/documented · calculation math on your inputs · StackSays rule our decision logic · StackSays estimate our judgement. The verdict cut-points are StackSays rules, not an industry threshold — there is no universal traffic number for these decisions.

Your situation

The recommendation updates as you answer. Share the URL to share it.

Corpus bigger than a context windowmore docs than fit in one promptContent changes frequentlydocs update daily/weeklyNeed source citationsshow which document an answer came fromCost per call mattersstuffing a huge context every call is expensiveSmall, stable doc seta handful of docs that rarely change

Recommendation

Probably yes

Based on the factors you set — 2 pushing toward adopting, 0 against.

Why yes

+22 Corpus exceeds context — Context windows are finite (per-model documented limits) — a corpus bigger than the window must be retrieved (RAG), not stuffedofficial fact
+12 Per-call cost — Stuffing a large context every call burns tokens; retrieving only relevant chunks is far cheaper at volumeStackSays rule

When to reassess

Use long-context while your docs fit the window and rarely change. Move to RAG when the corpus outgrows the window, updates often, or per-call token cost bites.

Minimal implementation

Start with long-context (stuff the docs) for a small stable set. Adopt RAG (chunk + embed + retrieve) when the corpus or cost forces it — see the vector-DB decision.

Risk of not doing it

Long-context on a big/changing corpus means truncated or stale answers and a large token bill; premature RAG means infrastructure you didn't need.

Cost & complexity once adopted

RAG adds an embedding + vector-store pipeline to build and maintain. Long-context adds tokens per call. Match to your corpus size and update rate.

This is a judgement tool, not a leaderboard. The StackSays rule and StackSays estimate factors are StackSays logic and opinion; the official fact factors link to official sources below. No industry-standard threshold is implied — the answer is driven by the conditions you set.

Official facts cited

Source	Verified
Long-context windows & limits (official)	2026-07-17

Only the factors tagged “official fact” rest on these sources; rules and estimates are StackSays logic and are labeled as such.

Next decision

Which vector database should I use? →Cheapest embedding API at your volume →