RAG & Retrieval Systems
The Problem
By hop three of a multi-hop retrieval, the intermediate query has drifted from the original question. The agent answers the intermediate query instead. Key terms shift definition between the source document and the retrieved chunk, and the embedding model encodes both as near-identical vectors.
How Ejentum Solves It
One API call forces your model to track semantic fingerprints across retrieval hops and detect the moment a key term shifts definition between the query and the retrieved content.
The Failures
- 01
The Pattern
Key terms shift definition between the original query and retrieved documents, causing the agent to answer a subtly different question than was asked
Why It Happens
Embedding models encode semantic similarity without preserving definitional boundaries, so a term that shifts meaning mid-pipeline produces near-identical vectors before and after the shift.
The Resolution
SI-015Semantic Drift DetectorTracks the semantic fingerprint of key terms across retrieval iterations, catching the moment a concept drifts from its original meaning.
Supported bySI-020 Identity Coherence Auditor - 02
The Pattern
Context window packed with the most semantically similar chunks, not the chunks that would most change the answer, burying signal in noise
Why It Happens
Embedding similarity is a proxy for relevance, not decision impact. The model cannot evaluate which chunks would change its output without already having generated it.
The Resolution
CA-034FalsificationistSeeks disconfirming evidence for every retrieval claim, testing whether removing a chunk would change the answer before committing context budget to it.
Supported bySI-041 Expected Value of Information Simulator - 03
The Pattern
In multi-hop retrieval, the intermediate query at hop 3 has drifted from the original question, and the agent answers the intermediate query instead
Why It Happens
Each query reformulation is an independent autoregressive generation optimizing local coherence with the previous step, not fidelity to the original information need.
The Resolution
SI-015Semantic Drift DetectorRe-anchors each hop to the original query fingerprint, catching the moment a reformulation drifts beyond acceptable semantic distance from the source question.
Supported byCA-001 Socratic Challenger
The Evidence
EjBench, 30 simulation tasks
Semantic drift is a focused, single-dimension problem. A single scaffold that anchors term definitions across hops outperforms multi-ability injection, which can introduce competing interpretations of the same term.
Task involved tracing cascade failure through a power grid. Baseline missed the second-order trip effect. Ki forced multi-step consequence propagation and caught that each generator failure re-concentrates load on survivors. Verification jumped from 0/3 to 3/3.
Run your next multi-hop retrieval through the API. See how the scaffold catches the query drift your embedding model misses.