← All Blogs
AI/ML · Knowledge Engineering

Conflict Detection in Large-Scale Knowledge Bases

When documents disagree — and they will — your system must know who wins and tell the user why. A systematic approach for any domain with 1000+ documents.

Mudit Kumar
Mudit Kumar · 5 min read

Why Conflicts Are Inevitable

In any organization with 1,000+ documents — legal, medical, financial, compliance, or technical — conflicts are guaranteed. Policies get updated but reference guides lag behind. Regional variations exist but aren't clearly scoped. Newer versions override older ones — but both remain indexed.

Without conflict detection, your RAG system picks one document at random (based on embedding similarity) and gives a confident wrong answer.

A doctor asks "What's the dosage for Drug X with renal impairment?" Two clinical guidelines disagree. A loan officer asks "What's the approval threshold?" Two policy versions conflict. The system must handle this — automatically.

Conflict Detection Architecture

CONFLICT DETECTION — END-TO-END FLOW INGESTION-TIME DETECTION 1. Extract atomic claims per section 2. Cluster claims by topic (embeddings) 3. Compare values within clusters 4. LLM pairwise comparison (flagged pairs) 5. Classify: factual|procedural|stale|scope 6. Auto-resolve (hierarchy rules) Resolved → Unresolved → Store winner in CONFLICTS table Flag for human review QUERY-TIME RESOLUTION User query → retrieve passages Lookup CONFLICTS for passage doc_ids Conflict found? RESOLVED Suppress loser Use winner only UNRESOLVED Show both to user with disclaimer Rerank with conflict annotations Generate answer (authority-first) feeds into
Fig 1: Conflict detection at ingestion time feeds pre-computed resolutions into query-time pipeline

Four Types of Conflicts

  1. Factual Contradiction — "Leave is 16 weeks" vs "Leave is 12 weeks" (High severity)
  2. Procedural Conflict — "Submit via portal" vs "Submit via email" (Medium)
  3. Staleness Gap — Guide references outdated policy version (Medium)
  4. Scope Ambiguity — Global vs regional rules unclear (Low)

Example Conflict

✓ Winner — Regional Policy 2025

"Parental leave entitlement is 16 weeks for all eligible employees, effective January 2025."

VS
✗ Outdated — Global Guide 2023

"Employees are entitled to 12 weeks of parental leave per the global framework."

Authority Hierarchy

When documents conflict, three deterministic rules resolve 90% of cases:

#1  Regional primary document (user's geography, latest)
 #2  Global primary document (latest version)
  #3  Regional reference/guide
   #4  Global reference/guide
    #5  Older version of anything

Rules:
• Primary ALWAYS beats reference (regardless of date)
• Regional ALWAYS beats global (for that region)
• Newer ALWAYS beats older (same document family)

Cost & Scale

Initial full scan (1,400 docs):  ~$25-40, 2-3 hours
Per new document (incremental):  ~$0.05, 30-90 seconds
Query-time lookup:               ~10ms (indexed table)
False positive rate:             <3% (with claim extraction)
Never hide unresolved conflicts. If you can't determine which source is correct, surface the ambiguity. Users prefer "these sources disagree" over a confident wrong answer.

Domain Applications

Takeaway

Detect early (ingestion time). Classify precisely (4 types). Resolve deterministically (hierarchy rules). Handle gracefully at runtime (suppress losers, surface ambiguity). Update incrementally (don't re-scan everything on each change).

Conflict DetectionNLPKnowledge BasesRAGDocument Intelligence
© 2026 Mudit Kumar. All content is intellectual property of the author. Unauthorized reproduction prohibited.