Concept Contrast Fable

Whether what is fed into the context is accurate and complete

Conclusion first:When RAG answers incorrectly, the problem may be that the material stuffed into the context is too mixed, or that the key material was never retrieved at all. Context precision penalizes irrelevant content, while context recall penalizes missing evidence.

When RAG answers incorrectly, the problem may be that the material stuffed into the context is too mixed, or that the key material was never retrieved at all. Context precision penalizes irrelevant content, while context recall penalizes missing evidence.

Fable

Testimony on the judge's desk and the absent witness

The county office was hearing a land case, and the clerk was responsible for delivering the case file to the judge's desk. The first time, afraid of missing evidence, he brought the genealogy, neighborhood gossip, and an unrelated ten-year-old lease. The judge read late into the night, and the real boundary-marker deed was buried under the pile of paper.

The second time, the advisor learned his lesson and sent only three sheets that seemed most important. The desk was clean, but a boundary-change letter was missed. The judge nearly ruled the land by the old boundary, and the neighboring village immediately cried foul.

The old judge had him distinguish two things. The materials on the desk must be clean, with few irrelevant papers mixed in, or judgment would be slowed by noise; the materials on the desk must also be complete, with no key testimony missing, or even clean materials would lead to a biased ruling.

The clerk later first listed the evidence that "must be present": the main contract, boundary change letters, boundary marker records, and the most recent tax register. Then he removed irrelevant gossip, duplicate copies, and personal letters that were only tangentially related. If a piece of material was uncertain, he wrote the reason in the margin instead of secretly stuffing it into a large pile.

At the next trial, the judge's desk held only a dozen or so pages, yet they covered all the key facts. The judge read quickly and missed fewer judgments.

The judge blamed the clerk for lacking judgment twice. The clerk felt wronged too: send too much and things get messy, send too little and something gets missed. The third time, he first laid out all possible evidence on the ground, then sorted it into piles by relevance to the dispute: boundary stones, letters changing the boundary, and witnesses from that year went in front; distant genealogy and gossip went behind.

The judge could finally hear cases based on the core evidence, and could also look up supplementary materials when needed. The more relevant papers on the desk, the sharper the judgment; but if one key paper is missing, no amount of tidiness helps. One test is how clean the materials submitted are; the other is whether the materials that should arrive actually do. The tension between the two keeps the case from being drowned by paper or led astray by omissions. From then on, the clerk asked two questions before submitting any file: has too much irrelevant material been mixed onto the desk, and is any paper that could truly change the verdict missing? Miss either question, and the judge will be misled.

When collecting the papers, the judge said that submitting materials is like preparing court evidence: accurate means sending fewer useless papers; complete means not missing a single vital paper. Seeking only accuracy will starve you to the bone; seeking only completeness will fatten you until you cannot see clearly. A good legal secretary must hold both lines at once.

Context precision

Context precision measures how much of the content fed into the model's context is actually relevant to the question. It focuses on quality degradation caused by noise, interference, and irrelevant evidence.

Context recall

Context recall measures how much of all the relevant content that should have been retrieved actually made it into the context. It focuses on incomplete answers caused by missing key evidence.

Story Mapping

Judge's desk: the context window given to the model
Irrelevant chitchat and old ledgers: noise caused by low context precision
The absent porter's testimony: low context recall leads to missing key evidence
Label whether each piece of paper is useful: evaluate context precision
The key testimony checklist on the wall: evaluating context recall requires a reference set of evidence that should have been retrieved

Applying it to RAG evaluation

When model citations are chaotic and answers are led astray by side branches, check context precision first; when answers look clean but miss conditions, exceptions, or timelines, check context recall first. In real projects, the two must be recorded separately, otherwise the team will mix retrieval recall, ranking, and context trimming problems together when fixing them.

Context precision Context recall