Context Grounding Score
0.92
True Evidence in Top Context
Hallucination Risk
Low (8%)
Context Distraction Probability
Retrieved Chunks
3 / 3
Ground Truth Coverage
Interactive 2D Vector Embedding Topology (D3.js)
Query
Ground Truth
Distraction Chunk
Retrieved Context Window
256 tokens/chunk
STATUS: Context Grounded
Dense-only similarity previously missed exclusion fine print; hybrid retrieval with cross-encoding brought liquid addendum into top-2 context window.
RECOMMENDED FIX: Increase BM25 sparse keyword weight (alpha=0.40) and set cross-encoder rerank threshold to 0.70 to eliminate standard warranty noise chunk c1.
Prompt Engineering Strategy (Flawed Approach)
"System: You are an expert. Strictly answer using ONLY provided context. Do NOT hallucinate!"
Result: If vector retrieval misses Chunk c2 (Exclusions), prompt instructions cannot force the LLM to synthesize un-retrieved facts. The model is forced to rely on c1 and hallucinate that liquid damage is covered under the standard 3-year term.
Vector Retrieval Optimization (True Solution)
"Hybrid Search (Alpha 0.4 BM25 / 0.6 Dense) + Reranker Cutoff 0.70"
Result: Calibrating hybrid retrieval explicitly surfaces Chunk c2 (Exclusions) and c3 (Liquid Addendum). Ground-truth evidence is guaranteed in context before LLM prompt generation begins.