Chain Reranker

When refining query results where relevance and diversity are critical to getting the best answers, relying on a single ranking strategy may not be flexible enough for some scenarios. The Vectara Chain Reranker addresses this challenge by enabling you to apply multiple ranking strategies sequentially, giving you absolute control over the ranking process.

This powerful capability enables you to completely customize Vectara's functionality to your specific needs, ensuring that your search results and generative AI applications align with your business requirements precisely.

Why use the chain reranker?

Chaining multiple rerankers is particularly useful in complex search scenarios that require several dimensions of ranking. For example, enhancing diversity with the Maximal Marginal Relevance (MMR) reranker, refining the results across multiple languages, and then boosting documents with specific metadata or other custom business logic with the User Defined Function reranker to finalize the results.

1. Multidimensional relevance: Combine multiple ranking strategies in sequence to optimize query results across various dimensions and meeting complex business scenarios.
2. Enhance RAG outcomes: Improve the retrieval process by sequentially applying multiple rerankers that balance relevance, diversity, and custom business priorities to produce more contextually appropriate AI-generated responses in Retrieval Augmented Generation (RAG) systems.
3. Absolute control: Tailor the ranking process to your exact specifications, ensuring that the query results reflect your custom business logic.

Enable the chain reranker

To enable the chain reranker, set the reranker object type to chain and then specify the multiple rerankers that you want to combine in a rerankers array. You can specify up to 50 rerankers in a chain.

Chain reranker examples

In this example, the Vectara Multilingual Reranker first refines the results with improved text scoring, while the User Defined Function reranker then boosts relevance based on metadata for popularity. This simple chain example can be useful for multilingual content platforms such as blogs, news aggregators, and knowledge bases.

{
  "reranker": {
    "type": "chain",
    "rerankers": [
      {
        "type": "customer_reranker",
        "reranker_name": "Rerank_Multilingual_v1"
      },
      {
        "type": "userfn",
        "user_function": "get('$.metadata.popularity') * get('$.metadata.score')"
      }
    ]
  }
}

In this example, the Vectara Multilingual reranker improves the precision of results by refining the output of initial models like Boomerang, while the User Defined Function reranker boosts relevance based on metadata for popularity. Now let's take a look at a more complex chain reranker example. In this e-commerce example, the Vectara Multilingual Reranker first refines the initial results, providing a better ranking order than Boomerang. This improves the overall relevance of the product results. The MMR reranker then diversifies the results so that the customer sees some variety, rather than only very similar products.

{
  "type": "mmr",
  "diversity_bias": 0.2
},

We set the diversity bias to 0.2 to have more emphasis on relevance while still providing some diversity. You can increase this value to 0.3 or 0.4 for more diversity.

Finally, we apply a User Defined Function that combines several e-commerce factors including a relevance score boost of 60%, a 20% boost if the product is currently in stock, and then another 20% boost that favors products with better customer reviews.

This chain reranker maintains a balance between showing relevant products that match what the customer is looking for, ensuring that the products are in stock, and also takes into account customer satisfaction rating of the products.

{
  "reranker": {
    "type": "chain",
    "rerankers": [
      {
        "type": "customer_reranker",
        "reranker_name": "Rerank_Multilingual_v1"
      },
      {
        "type": "mmr",
        "diversity_bias": 0.2
      },
      {
        "type": "userfn",
        "user_function": "get('$.metadata.relevance_score') * 0.6 + (get('$.metadata.in_stock') ? 0.2 : 0) + get('$.metadata.customer_rating') * 0.2"
      }
    ]
  }
}