FAQ and Q&A Matching

FAQ and Question-and-Answer (Q&A) matching leverages our neural retrieval capabilities to deliver precise answers and can help streamline support and knowledge management use cases. Enabling precise Q&A matching reduces the time users spend searching for information by enhancing both the accuracy and contextuality of responses. By parsing and understanding complex queries, Vectara transforms how organizations engage with their users, making interactions more context-aware.

Some users have frequently asked question (FAQ) databases or other forms of question databases where the use case demands that your users are trying to find the nearest question to their own, so you can provide them with the authoritative answer from the "answer" side of the question-answer database.

This approach may not offer the dynamic nature of Retrieval Augmented Generation (RAG), but it allows you to establish tight controls over the types of questions that users can ask and receive authorizative answers. These question-answer systems can be great for building RFP-answering systems for employees and FAQ lookups for customers.

Configure corpus for question matching

During corpus creation, set documents_are_questions to true to configures the corpus to use the query encoder for both indexing and querying. This is ideal for direct question-to-question matching, ensuring that the encoder used for indexing is aligned with the one used for querying, which improves match relevance.

We do not recommend changing the semantics setting to response for question matching. This method uses an encoder that is tailored for handling arbitrary textual content and would reverse the intended effect. It is often most effective when used in combination with a well-structured corpus and clear understanding of the user's search intent. Leave the default semantics setting.

Format data for question indexing

When you send data to Vectara for this use case, we recommend that you index the question in the title field and the answer to that question in the text content. For example:

document.json

{
  "id": "who-is-the-king-of-england",
  "type": "structured",
  "title": "Who is the King of England?",
  "sections": [
    {
      "text": "Charles III"
    }
  ]
}

Query for similar questions

Suppose you wanted to find the answer to a question related to this example. You can put Vectara into a document-matching mode by setting semantics to response. For example:

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

{
  "query": "Who's the English monarch?",
  "search": {
    "corpora": [
      {
        "corpus_key": "faq-corpus",
        "semantics": "response"
      }
    ],
    "offset": 0,
    "limit": 10
  }
}

This response setting disables Vectara's "question-answering" mode and instead tells it to find similar questions. This setting is useful when the objective is to discover content that is similar in context or subject matter to a given query

You can also add a filter expression of part.is_title = true to only match the questions.

Combine question matching and answering

Expanding on the previous example, we can help users find question or answer matches together by using querying multiple corpora. For example:

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

{
  "query": "Who's the English monarch?",
  "search": {
    "corpora": [
      {
        "corpus_key": "faq-corpus",
        "semantics": "response",
        "metadata_filter": "part.metadata.is_title = true"
      },
      {
        "corpus_key": "faq-corpus",
        "metadata_filter": "part.metadata.is_title IS NULL"
      }
    ],
    "offset": 0,
    "limit": 10
  }
}