Low-Precision Retrieval

Issue: Agent retrieves irrelevant chunks and synthesizes wrong answer.

Frequency: Occasional

Symptoms

• Citations do not support claim.
• Top-ranked results include documents that share keywords with the query but address a different topic entirely.
• Synthesized answer contains a claim that traces back to a semantically similar but factually unrelated chunk.
• Precision@5 drops noticeably below the precision@10 baseline for a specific query category (e.g., short, ambiguous queries).

Root Cause The embedding model conflates surface-level keyword overlap with true semantic relevance, so a chunk that shares vocabulary with the query but addresses an unrelated topic can score as similar as one that actually answers it. With no re-ranking stage running after initial dense retrieval to filter topically-similar-but-irrelevant chunks, and a similarity threshold set permissively enough to let borderline matches into the top-k, irrelevant chunks reach synthesis alongside relevant ones — and because no citation-grounding check verifies a chunk actually supports the specific claim before it’s used, the wrong chunk can end up cited as if it did.

Example

Query: "What's the policy on remote work for contractors?"
Vector search returns high-similarity chunks because "remote," "work," and "contractors"
appear frequently, but several top-5 results come from a "Remote Desktop Access" IT
security doc, not the HR contractor policy. The agent synthesizes an answer blending
VPN/remote-access rules with contractor status, producing a claim about "contractors
needing VPN approval for remote work" that neither source actually states.

Contributing Factors

• Embedding model conflates surface-level keyword overlap with true semantic relevance (e.g., “remote work” appearing in unrelated contexts).
• No re-ranking stage runs after initial dense retrieval to filter out topically-similar-but-irrelevant chunks.
• Retrieval similarity threshold is set too permissively, letting borderline-relevant chunks into the top-k.
• No citation-grounding check catches that a chunk doesn’t actually support the specific claim before it’s used in synthesis.

Eval Recipes

Test Cases

Metrics

Mitigation Strategies

Prevention

1. Learning-to-Rank Re-Ranker: After initial retrieval (recall-focused), apply ML re-ranking model to sort results by relevance. Train on query-document-relevance tuples from user feedback/annotations. Re-ranker produces relevance_score for each result; filter out low-scoring irrelevant chunks.
2. Citation Support Verification: Before synthesizing answer, verify each claim has supporting citation. Validate that cited chunk actually contains the claim. Block claims lacking citation support. Example: if claim=‘CEO announced X’, verify cited document contains ‘CEO announced X’.
3. Semantic Coherence Check: Compute semantic similarity between query and each retrieved chunk. Filter chunks below similarity threshold (e.g., < 0.6). Only use high-similarity chunks for synthesis.

Detection & Response

1. Precision@k Monitoring: Track precision@5, precision@10, precision@20 metrics in production. Estimate via user feedback (marked relevant/irrelevant). Alert if precision drops > 10% month-over-month.
2. Citation Verification Audit: Periodically audit synthesized answers. Verify each claim has supporting citation in retrieved chunks. Track: citation_accuracy_rate. Alert if < 95%.
3. Irrelevant Result Clustering: When user marks result irrelevant, analyze characteristics. Cluster irrelevant results to identify systematic false positive patterns (e.g., ‘all metaphor-based documents marked irrelevant’).

Architecture Patterns

1. Multi-Stage Ranking Funnel: Stage 1 (recall): broad vector search. Stage 2 (precision): re-rank with LTR model. Stage 3 (quality): LLM verification that top-k chunks support synthesis. Only use Stage 3 approved chunks.
2. Feature Engineering for Ranking: For each query-document pair, compute features: semantic_similarity, keyword_overlap, freshness, click_count, user_feedback_score. Feed to LTR model for ranking.
3. Citation Grounding Layer: Before synthesis, extract claims from synthesis. For each claim, find supporting citations in retrieved chunks. Only include claims with matched citations. Block ungrounded claims.

Metrics

1. precision_at_5_percent: Target: > 80%; Alert threshold: < 70%
2. precision_at_10_percent: Target: > 75%; Alert threshold: < 65%
3. citation_accuracy_rate_percent: Target: > 95%; Claims have supporting citations
4. irrelevant_in_top_10_rate_percent: Target: < 15%; Alert threshold: > 25%
5. ltr_model_ndcg_score: Target: > 0.85; Measure ranking quality

Alerts

1. Precision Degradation (P1 - Critical): Condition - precision@10 drops > 15% month-over-month. Action: Audit LTR model, check feature quality, potential retraining, ranking parameter tuning.
2. High Irrelevant Result Rate (P2 - Warning): Condition - > 25% of top-10 results marked irrelevant. Action: LTR model debugging, feature analysis, training data quality review.
3. Citation Grounding Failure (P1 - Critical): Condition - claim synthesized without supporting citation in retrieved chunks. Action: Block answer, escalate, audit synthesis logic, rerun retrieval.

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.