Cherry-Picking Evidence

Issue: Model Selectively Uses Supporting Evidence, Ignores Contradicting

Frequency: Common

Symptoms

• Answer presents one-sided view when context is balanced
• Caveats, exceptions, or limitations omitted
• User gets incomplete picture
• Decisions made on partial information

Root Cause Model may select evidence that best fits a coherent narrative, ignoring qualifying information or exceptions that would complicate the answer.

Example

Retrieved context:
"Clinical trials showed the treatment was effective in 67% of 
patients. However, 23% experienced significant side effects 
including nausea and headaches. The treatment is not recommended 
for patients with heart conditions or those over 65."

Query: "Is this treatment effective?"

Agent: "Yes, clinical trials showed the treatment was effective 
in 67% of patients."

Omitted: Side effects, contraindications, age restrictions

Result: User not informed of risks and limitations

Mitigation Strategies

1. Balanced response instructions: “Include caveats and limitations”
2. Structured extraction: Require pros/cons, conditions, exceptions
3. Comprehensiveness scoring: Measure coverage of context
4. Devil’s advocate check: Ask “what did you leave out?”
5. Multi-turn clarification: Follow up on omissions
6. Template responses: Ensure sections for limitations

Detection

• Compare answer coverage to context content
• Track caveat/limitation inclusion rates
• User feedback on incomplete answers
• Measure answer comprehensiveness

Mitigation Strategies

Prevention

1. Implement query-answer consistency validation: Decompose complex queries into atomic components and verify each component is addressed in the answer before returning. Use RAGAS Answer Relevancy metric (target: >0.75) with automatic re-generation for scores below threshold. Root cause mitigation: Prevents context-anchoring by explicitly binding answer generation to parsed query intent.

2. Apply multi-source consensus verification: Require answers synthesized from multiple sources to explicitly cite which sources support each claim and flag unresolved contradictions. Implementation: Use semantic similarity checks across source fragments to detect cherry-picked evidence patterns. Root cause: Ensures balanced representation of evidence when sources conflict.

3. Enforce comprehensive coverage checks: Implement structured extraction requiring explicit treatment of caveats, limitations, exceptions, and counterevidence for each claim type. Use template responses with mandatory caveat sections. Root cause: Prevents omission of qualifying information that would change user decision-making.

Detection & Response

1. Answer completeness monitoring: Measure coverage of query intents in generated answers. Track query decomposition rate (% of query components explicitly addressed) and flag responses with coverage <85%. Instrument RAG pipeline to log query-answer similarity scores per component. Alert on sustained scores <0.70.

2. Evidence balance scoring: For each answer, compute evidence distribution across sources and flag one-sided responses (>70% from single source on multi-source queries). Implement automated extraction of caveat/limitation mentions and track inclusion rates by query type. Target: >80% of medical/financial answers include relevant caveats.

Architecture Patterns

1. Query Intent Decomposition Graph: Parse complex queries into a DAG of atomic intents before retrieval. Each retrieved document is mapped to specific intent nodes. Answer generation must satisfy all leaf nodes. Validation layer computes coverage before response generation.

2. Evidence Consensus Engine: Maintain a fact graph where each claim is attributed to specific sources with confidence scores. Multi-source claims require consensus computation (intersection of sources supporting claim). Flagging layer surfaces contradictions to generation model.

3. Structured Response Templates: Use task-specific response schemas that enforce inclusion of: primary answer, supporting evidence, relevant caveats/exceptions, alternative interpretations, confidence bounds. Auto-flag template violations before user delivery.

Key Metrics

Alerts & Escalation

References

• Stanford Legal RAG Hallucinations - Selective evidence use
• Mindee: RAG Hallucinations Explained - Incomplete synthesis