Home/ Part VII — Multimodal & Long Context (Where AI Studio Gets Spicy)/22. Working With Documents and Large Text/22.4 Handling contradictions and multiple versions

22.4 Handling contradictions and multiple versions

Overview and links for this section of the guide.

On this page

Goal: detect and handle conflicts safely

Real document sets contain contradictions:

  • old docs vs new docs,
  • policy vs implementation,
  • different teams documenting the same behavior differently.

In long-context systems, contradictions cause the model to “average” sources and produce a confident but wrong synthesis. Your job is to surface conflicts explicitly.

Conflict-aware behavior is a feature

“I found conflicting sources” is the correct output in many cases. Don’t force a single answer when the sources disagree.

Why contradictions happen

Common causes:

  • Version drift: behavior changed but docs didn’t.
  • Different scopes: policy describes global behavior; implementation has environment exceptions.
  • Ambiguous language: “should” vs “must” vs “usually.”
  • Partial updates: one doc updated, another not.

A conflict-handling workflow

  1. Detect: require the model to list conflicting statements when they exist.
  2. Localize: show which chunks contain the conflict (with quotes).
  3. Classify: is it a version conflict, scope conflict, or ambiguity?
  4. Decide: apply a policy (see below) or ask a clarifying question.
  5. Record: add a note to a “contradiction ledger” so the conflict is handled consistently.

Decision policies (what to do when sources disagree)

Pick a policy and encode it in prompts:

  • Recency policy: prefer newer versions when version metadata is available.
  • Authority policy: prefer sources marked “canonical” (e.g., official spec) over informal docs.
  • Implementation policy: prefer what the code/tests do (if the codebase is the source of truth).
  • Escalation policy: when conflicts are high-impact, require human decision.

The important thing is consistency: the same conflict should not produce different answers across runs.

Copy-paste prompts

Prompt: answer with conflict detection

Answer the question using ONLY the provided chunks.

Rules:
- If sources conflict, do not guess. Report the conflict.
- Quote the conflicting statements and include chunk ids.
- If needed, ask 1 clarifying question to resolve the conflict.

Return JSON:
{
  "answer": string|null,
  "conflicts": [{
    "topic": string,
    "source_a": { "chunk_id": string, "quote": string },
    "source_b": { "chunk_id": string, "quote": string },
    "type": "version" | "scope" | "ambiguity" | "unknown",
    "suggested_resolution": string
  }],
  "follow_up_question": string|null
}

Where to go next