Context Selection — Why RAG Recall Does Not Equal Accuracy

Anyone who's shipped RAG has hit this number: 90% recall (top-k docs include the correct answer 90% of the time), but end-to-end accuracy is only 60%. Where did the 30% go?

Retrieval isn't the problem. Selection never happened.

Lost in the Middle — The 30% Empirical Gap

July 2023, Stanford's Liu and team published Lost in the Middle (arXiv:2307.03172), exposing a hidden flaw in every long-context LLM with one experiment:

Experiment design:

1. Multi-document QA task (20 docs + 1 question)
2. Only 1 doc contains the answer; 19 are distractors (topic-relevant, no answer)
3. Place answer-doc at different positions (1st / 5th / 10th / 15th / 20th)
4. Measure accuracy at each position

Result: every model tested (GPT-3.5, Claude, LLaMA) showed a U-shaped curve — peaks when answer sits at start or end, drops 30%+ when buried in the middle.

[] — see figure on page 3 of the original paper.

Of the 10 retrieved docs, content from doc 4 to doc 7 the model basically can't see. Even at 100% recall, the model can't use the part in the wrong position.

Not a bug. Joint product of transformer attention + training data distribution — start/end positions matter more in training data (paper abstracts, article conclusions). Anthropic reproduced in 200K models, Long context tips doc explicitly recommends putting key info at the end.

Three Root Causes — Not Just Position

Position bias is most visible, but RAG recall ≠ correct answer has three causes:

1. Position bias — Middle blindness

10 retrieved chunks, 5 relevant, sorted by relevance — relevance rank #1 isn't attention rank #1. Actively shove the most critical chunk to the end.

2. Distractor pollution

2024 Found in the Middle (He et al.) proves: more distractors in context (docs topic-similar but answer-empty), higher hallucination — the model "assembles" relevant nouns from distractors into fake answers.

Australian tax RAG retrieves 10 chunks — 3 GST, 7 income tax. User asks GST. The 7 income-tax chunks are distractors. Model says "GST is ..., note income tax has ..." — packing income tax details as if they were GST facts. Recalled right also recalled similar-but-wrong = pollution.

3. Instruction-following degradation

Anthropic's 200K context eval doc: longer context, lower adherence to system prompt hard rules (output format, prohibitions). At 50K with JSON output requirement, model occasionally drops fields. At 200K, field-dropping rate jumps noticeably.

Hidden cost of recalling more — you think "more retrieval is safer", really system instruction gets diluted.

Selection ≠ Retrieval

retrieval finds candidates, selection picks which actually enter context.

Intro RAG tutorials only teach retrieval — top-5 straight into prompt. Root cause of "90% recall, 60% accuracy".

Selection's Three-Stage Pipeline

Chapter 5 details each toolchain. Overall shape:

Stage 1 — Filter (coarse pass)

Drop by hard rules:

• Stale (2023 policy doc talking 2026 event)
• Source blacklist (recycled CSDN articles)
• Length anomalies (< 50 or > 5000 chars)

Cost dirt cheap (O(n)), drops 30-50% candidates.

Stage 2 — Rerank (fine pass)

Re-score with cross-encoder:

• bi-encoder (retrieval's) sees query or doc one at a time — fast, imprecise
• cross-encoder sees query + doc together, 30-50% more precise, 50-100x cost — only run on top-50

Mainstream: Cohere rerank-3 / BGE-reranker-v2 / Anthropic contextual retrieval.

Stage 3 — LLM-as-judge (final pass)

Cheap model (Haiku or 4o-mini) as judge:

• Input: query + single doc
• Output: does this answer the query? yes/no + reason

5-10x rerank cost, catches "topic-related but off-target" passages rerank lets through.

Three combined: retrieval 100 → filter 50 → rerank 10 → judge 3 → 3 enter context. Recall intact, distractor pollution to zero.

JR Real Case: daily-jobs Selection Pipeline

JR Academy runs 3 daily-jobs routines per day — one each for ai-essentials, ai-engineer, ai-engineer-rag bootcamp tracks, recommending 3 jobs per cohort. Textbook selection:

# tested: 2026-04-26 · routine: Daily Jobs - AI Engineer Bootcamp

Stage 0: 抓取（retrieval）
  WebFetch au.linkedin.com/jobs/junior-machine-learning-jobs
  WebFetch au.linkedin.com/jobs/ai-engineer-jobs
  → 抓回约 30 个 job

Stage 1: Filter（硬规则）
  剔除 recruiter spam（YO IT Consulting 这类同 job 多 location 群发）
  剔除 SEEK 来源（403 默认无效）
  剔除发布 > 7 天
  → 剩约 18 个 job

Stage 2: Rerank（按 tier 标签）
  打 3 个 tier 标签：
    aspirational（理想档 — Mid-Senior + T0/T1 brand）
    actionable（够得着档 — Junior/Graduate + AU Big 4）
    special（特殊机会 — Intern / Graduate Program / 2026-27 Start）
  → 每 tier 留 top 3-5 候选

Stage 3: LLM-as-judge（人称代入）
  以「JR Bootcamp 学员视角」给每个候选写 whyForLearners
  ≥ 30 字、必须具体、不允许模板填空
  → 每 tier 选 1 个最佳 → 共 3 个 final

最终 context 进 prompt：3 个 job 全文 + 3 个 whyForLearners

Not fancy algorithms — three filter stages plus one judge model. The 3 jobs that come out passed distractor screening + position ordering (tier is the position signal).

Students see "3 hand-picked jobs that matter to you", not "30 fetched, 3 chosen". Product value of selection.

High Recall + Weak Selection vs Low Recall + Strong Selection — Trade-off

JR experience: external production RAG goes high-recall + strong-selection. Internal tools go low-recall + weak-selection. Different latency, cost, quality bars.

Takeaway

Recall is retrieval's job. Accuracy is selection's job. Stuffing top-10 straight into prompt is a toy demo. Production RAG must split into retrieval / selection / composition. The middle stage (selection) is where the engineering work between 90% recall and 90% accuracy lives.

---

References

1. Liu et al. (2023-07-06). Lost in the Middle. arXiv:2307.03172.
2. He et al. (2024-03-08). Found in the Middle. arXiv:2403.04797.
3. Anthropic. Long context tips.
4. Anthropic. (2023-05-11). 100K context windows.
5. Liu. GitHub: lost-in-the-middle.

Production case: JR Academy omni-report Daily Jobs — retrieval / filter / rerank / LLM-judge three-stage pipeline.