Skip to main content

Hybrid Search

Semantic search enables users to discover information based not just on keywords, but on the contextual meaning and relevance of their queries. This neural search understands the intent behind user inquiries, leveraging natural language understanding to deliver more accurate and insightful search results.

Leveraging these semantic search capabilities, Vectara also provides a Hybrid Search option that offers a powerful and flexible approach to text retrieval. We combine partial, exact, and Boolean text matching with neural models which blends traditional, keyword-based search with semantic search in what is called a "hybrid" retrieval model.

For example, you can use this in Vectara to:

  • Include exact keyword matches for occasions where a search term was absent from Vectara's training data (e.g. product SKUs)
  • Disable neural retrieval entirely, and instead use exact term matching
  • Incorporate typical keyword modifiers like a NOT function, exact phrase matching, and wildcard prefixes of terms

Enable hybrid search by specifying lambda value at query time, specifically under the corpusKey. This value can range from 0.0 to 1.0 (inclusive).

As you ingest data and run queries, adjust the lambda value to achieve the perfect balance in answer quality.

      "corpusKey": [
        {
          "customerId": 123456789,
          "corpusId": 5,
          "semantics": 0,
          "metadataFilter": "",
          "lexicalInterpolationConfig": {
            "lambda": 0.025
          },
          "dim": []
        }

Experimenting with different lambda values​

The default value of lambda is 0, which disables exact and Boolean text matching. A value of 1 would disable neural retrieval instead, relying only on Boolean and exact text matching. Experimenting with the lambda value is useful if you're trying to evaluate how a keyword system like one based on Elasticsearch or Solr may compare to Vectara.

tip

💡 You can test queries with different lambda values in our API Playground and in the Vectara Console.

Vectara supports in-between values as well, which tells Vectara to try to consider both neural and Boolean and exact text matching and then to blend the scores of the results of the two different scoring models. Users often see best results by setting this lambda value somewhere between 0.01 and 0.1, and we typically recommend users start experimentation with a lambda value of 0.025.

Syntax Interpretation​

When interpreting query strings, Vectara treats the following syntax specially.

  • Words that are quoted must match exactly in that order. For example, the query blue shoes must match the word blue followed immediately by shoes.

  • A word fragment suffixed with an asterisk * is treated as a prefix match, meaning that it matches any word of which it is a prefix. For example, Miss* matches Mississippi.

  • Words prefixed with a minus - sign are excluded from the results. To extend on the previous example, -Mississippi would exclude results referencing the Magnolia State. Using -Miss* would exclude references to both Mississippi and Missouri.