Skip to main content

Question and Answering (QnA)

  Overview​

A question and answering (QnA) module does exactly what its name suggests. It will answer a question from you, based on the data present in the search results.

This is a more specific type of data transformation than one performed by a generative module. A QnA module looks to extract an answer to the question from the search data.

To this end, a QnA module may not provide an answer at all.

  Configuration​

  Enable modules​

To use QnA functionality, a qna-xxx module must be enabled in the Weaviate instance.

If you are using WCS, a qna module is enabled by default (see the documentation). Otherwise, you must configure your Weaviate instance to make sure that a qna-xxx module is enabled.

This is outside the scope of this unit, but you can refer to the module configuration for information on how to configure each module.

  Configure classes​

If only one qna module is enabled for the Weaviate instance, Weaviate will automatically use that module for all qna tasks.

On the other hand, if multiple qna modules are configured, you must define for each class which qna model to use, such as shown below.

{
  "classes": [
    {
      "class": "JeopardyQuestion",
      "moduleConfig": {
        "qna-openai": {
          "model": "text-davinci-002"
        }
      }
    }
  ]
}

Use of QnA functionality also requires that the target class be configured with a vectorizer (text2vec) module.

  About QnA queries​

  How it works​

Similarly to generative modules, a QnA search involves two steps, which are to:

  1. Perform a search; and then
  2. Attempt to extract an answer using the search results and the question.

Depending on whether a suitable answer was found, the answer sub-property under _additional properties may contain the answer. Some models will also return the position of the answer in the text.

If an answer is not found, the answer sub-property will indicate so.

  QnA syntax​

In a generative search, the same question text is used to both:

  • Perform the search, and
  • Extract the answer.

So, in the below example - the query "How many championships does Lewis Hamilton have?" is used to find the nearest object, from which Weaviate attempts to extract an answer to the question.

ask = {
  "question": "How many championships does Lewis Hamilton have?",
}

response = (
  client.query
  .get("WikiArticle", ["title", "_additional {answer {hasAnswer property result startPosition endPosition} }"])
  .with_ask(ask)
  .with_limit(1)
  .do()
)

print(json.dumps(response, indent=2))

In this query, the article includes an answer, and you can see that it has been extracted by the model.

See the JSON response
{
  "data": {
    "Get": {
      "WikiArticle": [
        {
          "_additional": {
            "answer": {
              "endPosition": 0,
              "hasAnswer": true,
              "property": "",
              "result": " Lewis Hamilton has seven World Drivers' Championship titles.",
              "startPosition": 0
            }
          },
          "title": "Lewis Hamilton"
        }
      ]
    }
  }
}

  How to specify properties​

You can specify the object properties in which the QnA module is to search through for the answer. Similarly to the case of a generative query, this may be useful if you want to reduce the length of the input, or be very specific about where the information should come from.

nearText search unaffected

Specifying properties to search only affects the answer extraction part of the query, as the underlying object vectors do not change.

ask = {
  "question": "How many championships does Lewis Hamilton have?",
  "properties": ["title"]
}

response = (
  client.query
  .get("WikiArticle", ["title", "_additional {answer {hasAnswer property result startPosition endPosition} }"])
  .with_ask(ask)
  .with_limit(1)
  .do()
)

print(json.dumps(response, indent=2))

And in this query, Weaviate is not able to answer the question as the required information is not available in the title field.

See the JSON response
{
  "data": {
    "Get": {
      "WikiArticle": [
        {
          "_additional": {
            "answer": {
              "endPosition": 0,
              "hasAnswer": true,
              "property": "",
              "result": " Lewis Hamilton has seven World Drivers' Championship titles.",
              "startPosition": 0
            }
          },
          "title": "Lewis Hamilton"
        }
      ]
    }
  }
}

  Object limits​

As this is also a two-step search, you can specify the number of objects for Weaviate to initially retrieve before attempting to extract the answer.

By setting the number of objects, you may increase the chance of retrieving the object that contains the answer to the specific question.

  Review​

  Key takeaways​

  • Question and Answer (QnA) search is another two-step search, which attempts to extract an answer to a specific question from the retrieved data before delivery.
  • To perform a QnA search, a qna-xxx module must be enabled in the Weaviate instance. The qna-openai module is enabled by default in WCS.
  • The QnA module will look for an answer in each retrieved object, returning the answer as an additional property.
  • If the QnA module does not identify an answer, it will indicate so in the response.