Enhancing Hybrid Search in MongoDB: Combining RRF, Thresholds, and Weights

In my previous blogs, I explored implementing basic hybrid search in MongoDB, combining vector and text search capabilities(https://dev.to/shannonlal/optimizing-mongodb-hybrid-search-with-reciprocal-rank-fusion-4p3h). While this approach worked, I encountered challenges in getting the most relevant results. This blog discusses three key improvements I've implemented: Reciprocal Rank Fusion (RRF), similarity thresholds, and search type weighting.

The Three Pillars of Enhanced Hybrid Search

1. Reciprocal Rank Fusion (RRF)

RRF is a technique that helps combine results from different search methods by considering their ranking positions. Instead of simply adding scores, RRF uses a formula that gives more weight to higher-ranked results while smoothing out score differences:

{
  $addFields: {
    vs_rrf_score: {
      $multiply: [
        0.4, // vectorWeight
        { $divide: [1.0, { $add: ['$rank', 60] }] },
      ],
    },
  },
}

2. Similarity Thresholds

To ensure quality results, I've added minimum thresholds for both vector and text search scores:

// Vector search threshold
{
  $match: {
    vectorScore: { $gte: 0.9 }
  }
}

// Text search threshold
{
  $match: {
    textScore: { $gte: 0.5 }
  }
}

This prevents low-quality matches from appearing in the results, even if they would have received a boost from the RRF calculation. In the example above I have chosen 0.9 for vector similarity score and 0.5 for text; however, you can adjust these based on your search results with your data.

3. Weighted Search Types

Different search types perform better for different queries. I've implemented weights to balance their contributions:

{
  $addFields: {
    combined_score: {
      $add: [
        { $multiply: [{ $ifNull: ['$vectorScore', 0] }, 0.4] },
        { $multiply: [{ $ifNull: ['$textScore', 0] }, 0.6] }
      ]
    }
  }
}

In this example I am giving a bit more weight to the text search results over the vector search, but again you can adjust these based on your search tests.

Putting It All Together

Here's a simplified version of the complete pipeline:

[
  // Vector Search with threshold
  {
    $vectorSearch: {
      index: 'ai_image_vector_description',
      path: 'descriptionValues',
      queryVector: embedding,
      filter: {
        userId: userId,
        deleted: false,
      }
    }
  },
  { $match: { vectorScore: { $gte: 0.9 } } },
  // RRF calculation for vector search
  {
    $group: {
      _id: null,
      docs: { $push: '$$ROOT' }
    }
  },
  // ... RRF calculation stages ...
  {
    $unionWith: {
      // Text search pipeline with similar structure
    }
  },
  // Final combination and sorting
  {
    $sort: { combined_score: -1 }
  }
]

Benefits and Results

This enhanced approach provides several benefits:

1. More relevant results by considering both ranking position and raw scores
2. Quality control through minimum thresholds
3. Flexible weighting to optimize for different use cases

The combination of these techniques has significantly improved our search results, particularly for queries where simple score addition wasn't providing optimal ordering.

Next Steps

Future improvements could include:

• Dynamic weight adjustment based on query characteristics
• Additional quality metrics beyond simple thresholds
• Performance optimization for larger datasets

By implementing these enhancements, we've created a more robust and reliable hybrid search system that better serves our users' needs.