Sentiment Bias in Modern Text Embedders Metric

Analysis of eleven encoders, including Qwen3-Embedding-4B, shows sentiment is highly decodable via linear probes (up to 91.3% accuracy) but absent from the cosine similarity metric. This dissociation means RAG and retrieval systems are effectively blind to sentiment information even when it is present in the embeddings. The gap exists due to pretraining objectives rather than prompting.