Inconsistent Answers

Issue: Agent gives different answers across turns.

Frequency: Occasional

Symptoms

• Contradictory responses for same fact/policy.
• Agent states a different policy value (return window, fee amount, SLA) later in the same conversation than it stated earlier, without acknowledging the change.
• Two independent sessions asking the identical policy question receive materially different answers, revealing retrieval or generation-level nondeterminism rather than a genuine policy change.

Root Cause There is no single canonical source per policy topic, so retrieval can surface a different, overlapping, or outdated document depending on how a question happens to be phrased, and the knowledge base itself often contains duplicate or conflicting entries for the same policy that were never deduplicated. Rather than being grounded in one fixed retrieved passage, the model reconstructs each answer from weak parametric memory on every turn, and because the agent keeps no structured record of facts it has already asserted earlier in the conversation, nothing checks a new answer against what was already said before it goes out.

Example

Agent (turn 2): "Our return window is 30 days from the delivery date."
...
User (turn 9): "So just to confirm, how many days do I have to return this?"
Agent (turn 10): "You have 14 days from the delivery date to return it."
User: "Wait, you told me 30 days earlier in this same chat."

Contributing Factors

• No single canonical source per policy topic — retrieval surfaces different, overlapping, or outdated documents depending on query phrasing.
• Model reconstructs the answer from weak parametric memory each turn instead of being grounded in a fixed retrieved passage or deterministic lookup.
• No conversation-level memory of facts already asserted, so later turns don’t check consistency against earlier ones before answering.
• Knowledge base contains duplicate or conflicting entries for the same policy that haven’t been deduplicated.

Eval Recipes

Test Cases

Metrics

Mitigation Strategies

Prevention

1. Single source-of-truth grounding: require every factual/policy claim to be generated from a retrieved passage (RAG) rather than parametric memory, and cite the source, since inconsistency across turns stems from the model reconstructing an answer from weakly-grounded internal knowledge each time rather than reading the same canonical source. Trade-off: adds retrieval latency and requires the knowledge base to be comprehensive and current, or grounding fails silently.
2. Conversation-level fact cache: once a fact/policy answer is given in a conversation, cache it and reuse the cached value for the rest of the session instead of re-deriving it on each turn. Trade-off: if the cached fact was wrong, the error now persists consistently through the whole conversation instead of surfacing as a visible contradiction.
3. Cross-turn consistency check before sending: before finalizing a response, compare the new claim against prior claims in the same conversation and recently retrieved passages, resolving contradictions before responding. Trade-off: adds an extra model call’s worth of latency and cost to every turn.

Detection & Response

1. Contradiction pair detection: run an automated same-conversation contradiction detector (NLI-style entailment check between turns) over live transcripts to catch conflicting answers to the same question. Response: flag the conversation for review and, if confirmed, send the user a corrected/clarifying follow-up.
2. Policy-drift monitoring across sessions: sample answers to a fixed set of canonical policy questions across many independent sessions and check variance; high variance indicates ungrounded or stale-retrieval-driven inconsistency. Response: investigate whether the knowledge base has duplicate/conflicting entries.
3. User-flagged contradiction reports: detect user language like “you just said…” and route to a contradiction-review queue. Response: human confirms and patches the knowledge base or retrieval ranking.

Architecture Patterns

1. RAG with canonical single-document retrieval: architect retrieval to prefer one canonical, deduplicated source per policy topic (versus multiple overlapping documents) so the same query reliably surfaces the same passage, removing a major source of turn-to-turn variance.
2. Conversation-memory checkpointing of asserted facts: maintain an explicit structured memory of facts/policies already asserted in the conversation and inject it into every subsequent generation call as a hard constraint, rather than relying on the model to recall its own prior turns from raw transcript context.
3. Deterministic policy-lookup service: move high-stakes, frequently-asked policy facts (pricing, refund windows, SLAs) out of free-form generation entirely into a deterministic lookup service the agent calls, eliminating generation-level inconsistency for the highest-impact facts.

Metrics

1. cross_turn_contradiction_rate: Target: <1% of multi-turn conversations; Alert on >2.5% weekly
2. canonical_policy_answer_variance: Target: <5% deviation across sampled sessions for fixed test questions; Alert on >15%
3. user_flagged_contradiction_rate: Target: <0.5% of conversations; Alert on >1.5%
4. grounded_response_rate: Target: >95% of factual/policy claims cite a retrieved source; Alert on <90%

Alerts

1. Contradiction Rate Spike (P2): Condition - cross_turn_contradiction_rate exceeds 2.5% over 24h. Action: sample flagged conversations, check for recent knowledge-base changes or retrieval regressions.
2. Canonical Answer Variance (P2): Condition - fixed-question eval shows >15% variance across sessions. Action: audit the knowledge base for duplicate/conflicting entries on the affected topic.
3. User-Reported Contradiction Surge (P3): Condition - user_flagged_contradiction_rate exceeds 1.5% weekly. Action: prioritize knowledge-base cleanup for the most-flagged topics.

Production Signals

Key Metrics

Alerts

References

• NIST-GenAI-Profile
• Note: Generative AI risks including confabulation, data privacy, information integrity, human-AI configuration, security, value chain.