Take fifty questions your system gets wrong. For each, check whether the passage containing the answer appeared anywhere in what was retrieved. That single split, one afternoon of work, gives you both numbers below and settles where the budget belongs.
| Scenario | End to end | Right | Wrong | What it tells you |
|---|---|---|---|---|
| Today | 65.3% | 326 | 174 | the number you have |
| If generation were perfect | 75.4% | 377 | 123 | the cap recall puts on you |
| If retrieval were perfect | 86.0% | 430 | 70 | the cap generation puts on you |
| Reranking the wider set | 77.1% | 386 | 114 | reachable without changing the index |
End-to-end accuracy, by recall and grounded accuracy
What a RAG system delivers overall, given how often it retrieves the right passage and how often it answers correctly once it has. Read across a row to see how little a better model buys when recall is low. The final column is a perfect model, which is the ceiling recall alone imposes.
| Recall | 70% grounded | 80% grounded | 90% grounded | 95% grounded | Perfect model |
|---|---|---|---|---|---|
| 50% | 40.0% | 45.0% | 50.0% | 52.5% | 55.0% |
| 60% | 46.0% | 52.0% | 58.0% | 61.0% | 64.0% |
| 70% | 52.0% | 59.0% | 66.0% | 69.5% | 73.0% |
| 80% | 58.0% | 66.0% | 74.0% | 78.0% | 82.0% |
| 90% | 64.0% | 73.0% | 82.0% | 86.5% | 91.0% |
| 95% | 67.0% | 76.5% | 86.0% | 90.8% | 95.5% |
Assumes the model answers correctly without the passage 10 percent of the time, which is a typical figure worth measuring rather than inheriting. Red is under 70 percent, amber under 85. Note the last column: at 70 percent recall a flawless model still delivers 73 percent.
A worked example
A support assistant over an internal policy corpus, reported at 65 percent end-to-end. The team's plan for the quarter is a better embedding model and a prompt rewrite, on the reasoning that the answers read as though the model is not understanding the material.
Fifty failed questions get split by hand in an afternoon. In 36 of them the passage containing the answer was never retrieved. Recall measures 72 percent; when the passage is present the answer is right 86 percent of the time. That puts end-to-end at 65 percent, which matches the reported number, and it means 71 percent of the errors are retrieval.
The ceiling is the part that changes the plan. Even a flawless model, on this index, reaches 75 percent. The entire prompt-and-model quarter is competing for at most ten points, and most of it is not available at all. Meanwhile recall at a wider candidate set is 88 percent, so a reranker over passages already being retrieved is worth about twelve points on its own, with no re-embedding and no migration.
The order that follows is recall first, reranking second, generation last. It is the opposite of the plan, and the afternoon that produced it cost less than the first week of the quarter would have.
The arithmetic, so you can check it
End-to-end accuracy is recall x grounded accuracy + (1 - recall) x accuracy without the passage. Set grounded accuracy to 1 and you have the ceiling recall imposes; set recall to 1 and you have the ceiling generation imposes. Neither can be exceeded by working on the other side.
The error split is the same expression rearranged: retrieval errors are (1 - recall) x (1 - accuracy without the passage) and generation errors are recall x (1 - grounded accuracy). Reranking headroom is the recall gained by widening k, multiplied by the accuracy difference that passage makes.
This treats retrieval as binary: the passage is in the set or it is not. Real systems have gradations that matter, including a passage retrieved but ranked so low it is truncated out of the context, and answers that need two passages where only one arrived. Both make your effective recall lower than the measured figure, so treat this as an optimistic bound. It is still the right bound to argue from, because it is already enough to settle which half the budget belongs in.
The reasoning is in recall is the silent killer and the split procedure in retrieval or generation, which half broke.
Questions
Recall is the share of questions where the passage containing the answer made it into the retrieved set at all. If it is not there, the model is being asked to answer from material that does not contain the answer, so no prompt, no larger model and no reranker recovers it. End-to-end accuracy cannot meaningfully exceed recall, which makes recall the ceiling on every downstream improvement.
Take fifty questions from real users that your system gets wrong, and for each one check whether the passage containing the answer appeared anywhere in what was retrieved. That is an afternoon of work and it splits your failures into the two buckets that need completely different fixes. Correct sources should be set by someone with the domain knowledge to judge them, not by the engineer and not by a model.
Whichever half holds more of your errors, which this computes. The common mistake is spending a quarter on prompting and model choice when most failures are questions where the passage was never retrieved. Improving the half that was already working produces no visible change and is the most expensive way to gather no information.
Only when recall rises appreciably as you widen the candidate set. That gap is the material reranking has to work with: passages already being retrieved but ranked too low. If recall is flat and low across k, the passages are not being found at all and a reranker has nothing to reorder.
Because it usually is not zero, and assuming zero understates your ceiling. Some questions the model answers correctly from what it already knew. That rate is worth measuring rather than guessing, and it is also the rate at which your system produces right answers for reasons your retrieval cannot explain or attribute.