Recommendation System

Many platforms struggle to keep users engaged with relevant content or products, often relying on simplistic matching algorithms. Vectara can be used as a semantic recommendation system out of the box in order to provide your users with semantically similar documents or products. This capability is useful when your goal is to increase engagement by accurately matching user interest or intent with relevant content.

Semantic recommendation system considerations

Before you begin using Vectara for a semantic recommendation system, it's useful to think through what types of recommendation flows you want to enable. For example:

• Do you want to recommend based on the entire document content or just 1 section/field like the document title?
• Do you want to recommend semantically similar content regardless of the source language or do you want to only match a particular language?
• Are you looking for exact duplicates or semantic similarity?

Exact duplicate matching

Exact duplicate matching can be useful when you want to ensure no duplicate content exists in your corpora or to find exact matches of "known bad" documents like those that might violate compliance rules in your organization. In general, we recommend that you use filter expressions for this.

Specifically:

1. When you index your content, hash your content using something like SHA-256 and add that as custom metadata on the document
2. To find similar content to a particular document, hash the entire document using the same hashing algorithm and then perform a filtered query to find exact hash matches

Similar document matching and near-duplicates

Sometimes, you want to recommend alternative products or web pages to a user that are similar to the one they're looking at or a recently purchased product. These use cases can be dealt with by using Vectara in a document-to-document search/recommendation platform. In order to do this, the most important change is that you'll need to use RESPONSE similarity measure (available to our Scale plan users). It's easier to explain how this is different by first explaining how the DEFAULT similarity works.

By default, Vectara is set up in a “question answering” mode. That is, Vectara's large language models are designed in principal to answer an end-user's question instead of finding similar documents when in their default mode. You can think of this as "the best answer to the question Who is the King of England? is not a document which has the text Who is the King of England? even though that has the highest overlap of keywords possible. Instead, Vectara is set up by default to be able to find the best answers to queries.

Document recommendation systems are not trying to answer questions though: they're trying to find the most similar documents. So for that use case, what you need to do is change the mode of the search to document similarity instead of question answering. You do that via the semantics key which is inside of the corpusKey block in the query.

If a user is looking at a document that has the text:

All about me

My name is Shane and I'm ...

and you wanted to find other documents that are similar to this, you can pass this document text to Vectara and set the semantics to RESPONSE. For example:

https://api.vectara.io/v2/query

{
  "query": "All about me\n\nMy name is Shane and I'm ...",
  "search": {
    "corpora": [
      {
        "corpus_key": "my-corpus",
        "semantics": "response"
      }
    ],
    "offset": 0,
    "limit": 10
  }
}

This will find documents that are most semantically similar to that document.

Further recommendation refinement

At times, it can be useful to further refine the recommendations. For example:

• Only suggest based on similar document titles
• Only suggest results that share the same language
• Only suggest results that were created by a particular user

In these cases, it can be useful to use Vectara's filter expressions. There are out of the box filters for title and language and you can make use of additional metadata you add, such as the author or publication date.