Knee Reranking

Retrieval Augmented Generation (RAG) systems often struggle with determining optimal cutoff points for query results. Fixed score thresholds provide a simple solution but lack the flexibility to adapt to varying score distributions across queries. Knee reranking addresses this challenge by detecting natural boundaries between relevant and irrelevant results automatically.

Knee reranking combines statistical analysis with configurable parameters to provide intelligent, adaptive filtering. Designed specifically to work after the Slingshot reranker, it analyzes score patterns to identify significant drops in relevance while maintaining safeguards against over-aggressive filtering. For more details about how this reranker works, see this blog post.

Enable knee reranking

Enable knee reranking by adding it your reranking chain after the Slingshot reranker. The default settings balance precision and recall, making them suitable for most use cases.

Default Configuration

{
  "reranker": {
    "type": "chain",
    "rerankers": [
      { "type": "customer_reranker", "reranker_name": "Rerank_Multilingual_v1" },
      { "type": "userfn", "user_function": "knee()", "cutoff": 0.5 }
    ]
  }
}

Customize the behavior of knee reranking through two key parameters:

• Sensitivity: Controls how sharply the score must drop to identify a cutoff. Use higher values when you need increased precision.
  • Default: 0.5
  • Range: 0 to 1
• Early bias: Adjusts preference for cutting off results earlier in the ranking list. Use higher values when you want to reduce latency and focus on top results.
  • Default: 0.2
  • Range: 0 to 1

Combining parameters

A high sensitivity combined with high early bias narrows the results to only the top-scoring entries. Alternatively, a low sensitivity combined with low early bias broadens the results, favoring inclusivity.

• High Sensitivity: Identifies subtle score drops, triggering knee reranking and providing stricter relevance filtering.
• High Early Bias: Places stronger emphasis on earlier knees if they are similar in magnitude
• Low Sensitivity and Early Bias: Captures broader sets of results, suitable for exploratory or discovery tasks.

Test different parameter values to determine the best fit for your use case.