Skip to main content
Search & Retrieval

Source Conflict Resolution

Source conflict resolution is how a search or RAG system detects and handles contradictory sources, prioritizing controlling authority and making uncertainty explicit.

Also known as: Conflict handling, Contradiction management, Source disagreement resolution

Definition

Source conflict resolution is the set of rules and workflows a retrieval or RAG system uses when relevant sources disagree. In legal settings, conflicts can be real (different rules across time or jurisdictions) or apparent (different interpretations, missing context, or superseded versions).

Why it matters

  • Correctness: legal answers must follow controlling authority and applicability.
  • Safety: hiding conflicts can create false certainty and compliance risk.
  • Trust: users value transparency when sources do not align.
  • Better triage: routes genuinely ambiguous cases to human review.

How it works

Conflict resolution is usually a combination of detection + policy:

Retrieve -> detect conflicts -> apply hierarchy + applicability -> present outcome + citations -> escalate if ambiguous

Common strategies:

  • Prefer sources with higher authority and clearer applicability (jurisdiction, dates)
  • Distinguish “superseded” vs “co-existing” rules (e.g., different periods)
  • Show both positions with explicit conditions, not a single blended statement
  • Require human oversight for high-impact or unresolved conflicts

Practical example

Two sources state different withholding tax rates. Investigation shows one applies before an amendment date. The system highlights the effective dates, cites both versions, and uses the in-force text for the requested period.

Common questions

Q: Can an LLM resolve conflicts by itself?

A: It can summarize and compare, but the system should enforce authority/applicability policies and escalate when uncertainty remains.

Q: Should conflicts always be shown to users?

A: For legal and tax, it is usually safer to show conflicts (or at least signal uncertainty) rather than hiding them behind a single confident answer.

References

Regulation (EU) 2024/1689 (EU AI Act).