NemoClaw Knowledge Wiki

vanilla-rag

2 min read

🗂️ AI & Agents · View mindmap

Vanilla Rag

Vanilla Rag is a context engineering technique that improves Retrieval Augmented Generation (RAG) systems by reducing hallucination through selective filtering of retrieved documents. RAG systems augment language models with external knowledge by retrieving relevant passages before generation, but the retrieved content often includes irrelevant or contradictory passages that can degrade output quality. Vanilla Rag addresses this by applying re-ranking and pruning operations to the retrieved document set, retaining only the most relevant passages while discarding those likely to introduce noise or errors.

Mechanism

The technique operates in two stages. First, re-ranking algorithms assess the relevance of each retrieved passage relative to the user query, reordering them by quality. Second, pruning removes passages below a relevance threshold or those that may conflict with higher-confidence content. This dual approach reduces the likelihood that a language model will encounter contradictory information or peripheral details that could trigger confabulation, while maintaining access to genuinely useful context.

Applications and Limitations

Vanilla Rag is particularly effective in knowledge-intensive tasks where retrieval quality directly impacts answer accuracy. However, the approach depends heavily on the quality of the re-ranking mechanism and the selection of pruning thresholds. Overly aggressive pruning may remove useful context, while conservative settings may fail to eliminate problematic passages. The technique is most practical in systems where computational cost for additional ranking steps is acceptable and where ground truth relevance can be reasonably approximated.

Source Notes

Graph View

Backlinks