Detecting Path Contradictions in Reasoners: The Truth Auditor

In a large knowledge graph, you will eventually find Conflicting Truths.

• Path A: (Doc 1) "The project ends in June."
• Path B: (Doc 2) "The project was extended to August."

If you just feed both to an LLM, it might get confused or "Average" them out (e.g., "The project ends in July"). A professional Graph RAG system must be a Truth Auditor. It must identify when two paths contradict each other and apply a logic for Conflict Resolution.

In this lesson, we will look at Contradiction Detection. We will learn how to use the "Metadata" of the graph (Timestamps, Author PageRank, Source Trust) to decide which path is the "Winner." We will see how to build a prompt that forces the AI to highlight conflicts rather than hiding them.

---

1. The Contradiction Matrix

When an AI finds two conflicting facts, it should look at:

1. Freshness: Path B is from 2025; Path A is from 2024. -> B wins.
2. Authority: Path A is from the Official_Policy node; Path B is from a Slack_Chat node. -> A wins.
3. Frequency: Path A is supported by 5 different documents; Path B is supported by 1. -> A wins.

---

2. Prompting for Conflict Awareness

Don't let the AI decide quietly. Force it to be an "Honest Reporter."

INSTRUCTION:
If you find two paths that state conflicting information for the same attribute:
1. List both paths and their sources.
2. Identify which source is more authoritative (e.g., PDF vs Chat).
3. Identify which source is more recent.
4. State your conclusion and WHY you chose it.

---

3. The "Cross-Verification" Hop

If the AI finds a conflict, it should perform an additional Recursive Search (Lesson 3) to find a third "Tie-breaker" fact in the graph.

• "I've found two different dates. Let me search for the 'Project Manager' node to see if they have a definitive statement on the deadline."

graph TD
    User --> Q[Question]
    Q --> P1[Path A: June]
    Q --> P2[Path B: August]
    
    subgraph "The Resolution Engine"
    P1 --- V{Judge}
    P2 --- V
    T[Tie-Breaker Path] --- V
    end
    
    V -->|Verdict: August| A[Final Answer]
    
    style V fill:#f4b400,color:#fff
    style A fill:#34A853,color:#fff

---

4. Implementation: A Conflict Resolution Prompt

EVIDENCE 1: (ProjectX)-[:DEADLINE]->(June) [Source: Meeting Notes, Date: 2024-01]
EVIDENCE 2: (ProjectX)-[:DEADLINE]->(August) [Source: Official Jira, Date: 2024-05]

TASK: Find the final deadline.
REASONING: Evidence 2 is more recent (May vs Jan) and from a more authoritative source (Official Jira vs Meeting Notes).
ANSWER: The deadline is August.

---

5. Summary and Exercises

The Truth Auditor ensures that your AI isn't "Easily Misled."

• Metadata (Time/Source) is the primary tool for conflict resolution.
• Transparency about conflicts builds user trust.
• Tie-breaker searches provide a biological-like "Sense Check."
• Authority Weighting prevents "Social Noise" (Chat) from overriding "Formal Knowledge" (Docs).

Exercises

1. Conflict Resolution: You have a "Public Wiki" that says a product is $50 and a "Private Catalog" that says it's $60. Which one should the AI tell a customer? Which one should it tell a salesperson?
2. The "Tie-Breaker" Search: If two employees claim to be the "Lead" of the same project, what and where in the graph would you search to find the "Mathematical Truth"? (Hint: HR node, Payroll node, or Project Ownership node).
3. Visualization: Draw two paths going to the same node. Give each path a "Weight" score. Cross out the path with the lower score.

In the final lesson of this module, we will verify the result: The 'Graph-Grounded' Answer: Verifying the logic.