Reranking

Reranking adjusts the order of search results using additional signals beyond initial relevance scores, such as user feedback and popularity metrics.

​

Understanding Reranking

​

Key Concepts

1. Ranking Signals

   • Initial relevance score
   • User feedback
   • Popularity metrics
   • Custom weights

2. Weight Configuration

   • Signal importance
   • Score normalization
   • Dynamic adjustment
   • Feedback incorporation

​

Search Interactions

​

Interaction Types

1. Positive Signals

   {
     "feature_id": "prod_123",
     "interaction_type": "click",
     "position": 2,
     "metadata": {
       "duration_ms": 45000,
       "device": "mobile"
     }
   }
   

2. Negative Signals

   {
     "feature_id": "prod_456",
     "interaction_type": "skip",
     "position": 1,
     "metadata": {
       "interaction_reason": "not_relevant"
     }
   }
   

​

Recording Interactions

1. Required Fields

   • feature_id: Unique identifier for the result
   • interaction_type: Type of interaction (click, skip, etc.)
   • position: Position in search results
   • metadata: Additional context (optional)

2. Session Tracking

   {
     "session_id": "sess_abc123",
     "search_request": {
       "queries": [
         {
           "type": "text",
           "value": "blue dress",
           "vector_index": "text"
         }
       ]
     }
   }
   

​

Reranking Configuration

​

Basic Structure

{
  "reranking_options": {
    "weights": {
      "feedback": 0.7,
      "popularity": 0.3
    },
    "enable_reranking": true,
    "interaction_window_days": 30,
    "min_interaction_threshold": 5
  }
}

​

Advanced Options

1. Interaction Windows

   {
     "reranking_options": {
       "interaction_weights": {
         "recent": 0.6,
         "historical": 0.4
       },
       "time_decay": {
         "half_life_days": 7
       }
     }
   }
   

2. Signal Thresholds

   {
     "reranking_options": {
       "min_signals": {
         "clicks": 10,
         "views": 50
       },
       "confidence_threshold": 0.8
     }
   }
   

​

Implementation Guide

​

Recording Interactions

1. Client-Side Integration

   // Track click interaction
   await mixpeek.trackInteraction({
     feature_id: resultId,
     interaction_type: "click",
     position: resultIndex,
     metadata: {
       device: "mobile",
       duration_ms: viewDuration,
     },
   });
   

2. Batch Processing

   # Process multiple interactions
   interactions = [
     {
       "feature_id": "prod_123",
       "interaction_type": "click",
       "position": 2
     },
     {
       "feature_id": "prod_456",
       "interaction_type": "skip",
       "position": 1
     }
   ]
   
   await mixpeek.bulk_track_interactions(interactions)
   

​

Enabling Reranking

# Configure search with reranking
search_request = {
  "queries": [{
    "type": "text",
    "value": "summer dress",
    "vector_index": "text"
  }],
  "reranking_options": {
    "weights": {
      "feedback": 0.7,
      "popularity": 0.3
    },
    "enable_reranking": true
  }
}

results = await mixpeek.search(search_request)

​

Best Practices

​

Optimization Tips

1. Weight Tuning

   • Start with balanced weights (0.5/0.5)
   • Adjust based on interaction volume
   • Monitor impact on key metrics
   • A/B test different configurations

2. Data Quality

   • Filter bot/spam interactions
   • Validate interaction data
   • Handle edge cases
   • Maintain consistent tracking

​

Error Handling

​

Common Issues

1. Invalid Configuration

   {
     "error": "invalid_weights",
     "message": "Weights must sum to 1.0",
     "weights": {
       "feedback": 0.7,
       "popularity": 0.5
     }
   }
   

2. Resolution Steps

   • Validate weight configurations
   • Check interaction data format
   • Monitor reranking performance
   • Handle missing signals gracefully

For implementation details and examples, see the Interactions API Reference.