Wei-Cheng Chiu

邱偉誠

Taipei, Taiwan · UTC+8

making every FLOP count

Blog

The 0% RAG result failed the harder test

On this page ⌄

1. The two evaluations were testing different problems｜兩次評估測的是不同問題
2. Define the failure before reporting the percentage｜回報百分比前先定義失敗
3. Same-model validation shares the generator’s blind spots｜同模型驗證會共享 generator 的盲點
4. Better retrieval did not make abstention safer｜更好的 retrieval 沒有讓拒答更安全
5. What I would deploy from these results｜根據這些結果，我會怎麼部署

The 0% RAG result failed the harder test｜RAG 的 0% 結果沒有通過更難的測試

2026-05-31 · updated 2026-07-21 · RAG / evaluation / grounding

Grounded RAG reference architecture · Grounded RAG 參考架構 · Source: Microsoft Learn

My first evaluation made a tidy claim: a guarded prompt cut hallucination from 4/10 to 0/10 on unanswerable questions. The run was real, reproducible, and almost useless as a headline. Ten questions were too few, and “out of corpus” was the easy failure mode—the model had no relevant passage tempting it toward a plausible answer.¹

第一次評估得到一個很整齊的結論：面對無法回答的問題，加上 guard prompt 後，hallucination 從 4/10 降到 0/10。結果是真的，也能重現，但幾乎不能拿來下標題。十題太少，而且「語料庫外」是容易的失敗模式；模型根本找不到相關段落，自然少了被誘導出合理錯答的機會。¹

The scaled rerun used 100 answerable and 100 adversarially unanswerable SQuAD 2.0 questions. Each near-miss question retrieves an on-topic passage that does not contain the answer. Here the guarded generator hallucinated on 48/100 questions; without the guard it failed on 78/100.² The 0% result did not generalize. That reversal is the useful result.

擴大後的重跑使用 100 題可回答、100 題經對抗設計而無法回答的 SQuAD 2.0 問題。每個 near-miss 問題都會取回主題相關、卻不含答案的段落。這次有 guard 的 generator 在 48/100 題出現 hallucination，沒有 guard 時是 78/100。² 原本的 0% 無法泛化；這次反轉才是有用的結果。

The two evaluations were testing different problems｜兩次評估測的是不同問題

Evaluation／評估	Unanswerable set／無法回答資料集	Guarded／有 guard	Unguarded／無 guard	What it measures／測量內容
Initial harness／初始 harness	10 out-of-corpus questions／10 題語料庫外問題	0/10	4/10	Can the model abstain when evidence is plainly absent?／證據明顯不存在時，模型能否拒答？
Scaled rerun／擴大重跑	100 adversarial near-misses／100 題對抗式 near-miss	48/100	78/100	Can it resist a relevant-looking passage that lacks the answer?／面對看似相關卻沒有答案的段落時，模型能否避免錯答？

The first result still describes those ten examples. It just cannot support “this guardrail solves hallucination.” The harder set changes both the point estimate and the mechanism: the context is no longer empty or irrelevant, so “answer only from context” can steer the model toward a wrong span instead of toward abstention.

第一個結果仍能描述那十題，卻不足以支持「這個 guardrail 解決了 hallucination」。較難的資料集同時改變估計值與失敗機制：context 不再是空白或不相關內容，「只能根據 context 回答」反而可能把模型推向錯誤 span，而不是拒答。

Define the failure before reporting the percentage｜回報百分比前先定義失敗

For the unanswerable slice, let \(h_i=1\) when answer \(i\) contains a claim unsupported by the retrieved passage. The reported hallucination rate is

在無法回答的資料切片中，若答案 \(i\) 包含取回段落無法支持的主張，就令 \(h_i=1\)。Hallucination rate 定義如下：

\[ \widehat{H}_{\text{unanswerable}} = \frac{1}{N_{\text{unanswerable}}} \sum_{i=1}^{N_{\text{unanswerable}}} h_i. \tag{1} \]

That denominator must travel with the percentage. A zero from \(N=10\) and a zero from \(N=1{,}000\) are not the same amount of evidence. The scaled report therefore includes Wilson 95% confidence intervals and paired exact McNemar tests where two gates score the same answers.²

百分比必須連同分母一起看。\(N=10\) 的零和 \(N=1{,}000\) 的零，證據強度完全不同。因此，擴大後的報告加入 Wilson 95% confidence intervals；兩個 gates 評分同一批答案時，也使用 paired exact McNemar tests。²

Retrieval recall and hallucination rate also stay separate. An answer may retrieve the gold passage and still invent a claim; an unanswerable question may retrieve the intended near-miss perfectly and still require abstention.

Retrieval recall 與 hallucination rate 也要分開。即使取回 gold passage，答案仍可能捏造主張；無法回答的問題即使精準取回預定的 near-miss，仍然必須拒答。

Same-model validation shares the generator’s blind spots｜同模型驗證會共享 generator 的盲點

The original validate() step asked the same 8B model family to judge its own unguarded answers. It caught only one of four hallucinations in the small run. At scale, the plain self-judge reached 27% recall. Switching to an independent 8B model family raised recall to 41%, but 53 of 96 hallucinations escaped both judges.³

原本的 validate() 讓同一個 8B model family 判斷自己的 unguarded answers。小型實驗的四次 hallucination 只抓到一次；擴大後，plain self-judge 的 recall 是 27%。換成獨立的 8B model family 後，recall 升到 41%，但 96 次 hallucination 中仍有 53 次同時逃過兩個 judges。³

Gate on the same 200 answers／同一批 200 答案的 gate	Precision	Recall	F1
Same-family judge／同 family judge	79%	27%	0.40
Cross-family judge／跨 family judge	74%	41%	0.52
Same-family judge + chain of thought／同 family + CoT	81%	40%	0.53
Cross-family judge + chain of thought／跨 family + CoT	62%	49%	0.55
MiniCheck-FT5 770M	66%	36%	0.47
Cross-family judge OR MiniCheck／跨 family judge 或 MiniCheck	63%	55%	0.59

The smallest model in that table gives the most interesting signal. MiniCheck is a 770M grounded NLI verifier: it compares the answer against evidence instead of deciding from its parametric memory. It recovered 12 of the 53 errors missed by both 8B judges—more blind spots than any generative-judge variant.⁵³

表中最小的模型反而提供了最有意思的訊號。MiniCheck 是一個 770M grounded NLI verifier，會把答案和證據直接比對，不靠 parametric memory 判斷。兩個 8B judges 都漏掉的 53 個錯誤中，它補抓到 12 個，比任何 generative-judge variant 都多。⁵³

This is why “use a bigger judge” is not the default answer. A separate 70B experiment found no meaningful recall advantage over the cross-family 8B judge on its paired set, while the grounded verifier continued to catch errors the large parametric judge missed.¹

因此，不能預設換成更大的 judge 就會改善結果。另一組 70B 實驗在 paired set 上，recall 並未明顯優於 cross-family 8B judge；grounded verifier 則繼續抓到大型 parametric judge 漏掉的錯誤。¹

Better retrieval did not make abstention safer｜更好的 retrieval 沒有讓拒答更安全

Adding a cross-encoder reranker improved retrieval recall@3 from 85% to 96% and substring answer accuracy from 70% to 81%. Hallucination did not improve: guarded answers moved from 48% to 52%, and unguarded answers from 78% to 86%; neither change was statistically significant.⁴

加入 cross-encoder reranker 後，retrieval recall@3 從 85% 升到 96%，substring answer accuracy 從 70% 升到 81%。Hallucination 沒有改善：guarded answers 從 48% 變成 52%，unguarded answers 從 78% 變成 86%；兩項變化都沒有統計顯著性。⁴

That result is easy to misread. The reranker worked—it surfaced more relevant passages. On an adversarially unanswerable question, however, a more relevant answerless passage gives the generator a better near-miss. Retrieval quality and abstention safety are different axes.

這個結果很容易被看反。Reranker 確實找到了更多相關段落；但對經對抗設計且無法回答的問題而言，更相關但沒有答案的段落，會提供更像答案的 near-miss。Retrieval quality 與 abstention safety 是兩條不同的軸。

What I would deploy from these results｜根據這些結果，我會怎麼部署

The measured gate with the best coverage is the union of two different failure detectors: block when either the cross-family judge or MiniCheck flags the answer. It reaches 55% recall on the common Phase-5 run, with 63% precision. That is not a solved system; 29 of 96 hallucinations still escaped every tested gate.

實測 coverage 最好的 gate 是兩種 failure detectors 的聯集：cross-family judge 或 MiniCheck 任一標記答案就阻擋。在共同的 Phase-5 run 上，recall 是 55%，precision 是 63%。這套系統仍未解決問題；96 次 hallucination 中，還有 29 次逃過所有測試過的 gates。

So the production contract is explicit:

因此，production contract 要明確寫成以下五點：

1. Generate with an abstention instruction, because it still cuts the hard-set failure rate by 30 points.
   生成時加入 abstention instruction；它仍能讓 hard-set failure rate 降低 30 個百分點。
2. Verify against evidence, not only with another generative model.
   直接對照證據驗證，不能只加另一個 generative model。
3. Log the rejected and escaped cases with retrieved spans so the next evaluation set is built from real failures.
   連同 retrieved spans 記錄被擋下與逃脫的案例，讓下一版 evaluation set 來自真實失敗。
4. Report precision and recall, not just the residual hallucination rate; an aggressive gate can look safe by blocking correct answers.
   同時回報 precision 與 recall；只看 residual hallucination rate，可能把大量阻擋正確答案的 gate 誤認為安全。
5. Keep a deterministic adversarial set and rerun paired comparisons whenever the model, retriever, prompt, or judge changes.
   保留 deterministic adversarial set；model、retriever、prompt 或 judge 有變動時，重新執行 paired comparisons。

The adversarial set changed the claim: the prompt guard reduced hallucination but did not eliminate it. Future evaluations should include relevant-looking passages that still lack the answer.

對抗資料集改寫了原本的結論：prompt guard 能降低 hallucination，卻無法消除它。後續評估必須加入看似相關、實際上仍不含答案的段落。

→ Reproduce the runs, inspect row-level verdicts, and read the statistical caveats:
重現實驗、檢查逐列 verdict，並閱讀統計限制： github.com/waynehacking8/nim-agent-blueprint

---

1. NIM Agent Blueprint, the primary repository containing the agent graph, datasets, committed row-level verdicts, reports, and reproduction commands.／主要 repo，包含 agent graph、資料集、逐列 verdict、報告與重現命令。 ↩↩↩↩

2. SQuAD 2.0 evaluation report, including the N=200 design, Wilson intervals, and paired comparisons.／包含 N=200 設計、Wilson intervals 與 paired comparisons。 ↩↩↩↩

3. Judge-variant report, comparing same-family, cross-family, chain-of-thought, MiniCheck, and union gates on shared answers.／比較同 family、跨 family、CoT、MiniCheck 與 union gates。 ↩↩↩↩

4. Reranker ablation, with paired retrieval, accuracy, and hallucination results.／包含 paired retrieval、accuracy 與 hallucination 結果。 ↩↩

5. MiniCheck: Efficient Fact-Checking of LLMs on Grounding Documents, the grounded verifier used in the evaluation.／評估使用的 grounded verifier。 ↩↩