Compress evidence extraction to reduce token usage by 15-20%

[apenab]

Problem

Evidence extraction currently pulls near-full page chunks (~2000-4000 chars), which inflates token usage.

Proposal

Compress evidence extraction to include only the relevant section plus minimal surrounding context (~800-1200 chars), while preserving traceable grounding.

Expected Impact

• Estimated token reduction: ~15-20%
• Example target: 2.2M -> 1.8M tokens (~-18%)

Risk

• Moderate: missing key context may hurt grounding if extraction is too aggressive.

Implementation Considerations

• Keep extraction deterministic and auditable (store source span references).
• Include configurable context window around matched evidence.
• Add fallback to larger context when confidence is low.

Acceptance Criteria

• Evidence payload size is reduced to target range for most samples.
• Grounding metrics do not regress beyond agreed tolerance.
• Benchmark report includes before/after token use and quality.
• Failure analysis documents any context-loss errors.