This is a utility node that simply merges multiple lists of answers to produce a single list. It can be useful if you're using multiple PromptNodes, each generating an answer, followed by a node that accepts a single answer rather than multiple lists of answers. You can then use JoinAnswers after the PromptNodes.

Basic Information

• Pipeline type: Used in query pipelines.
• Nodes that can precede it in a pipeline: AnswerDeduplication, PromptNode with AnswerParser to make sure it outputs Answer, Reader, ReferencePredictor
• Nodes that can follow it in a pipeline: AnswerDeduplication, ReferencePredictor
• Node input: Answers
• Node output:Answer
• Available node classes: JoinAnswers

Join Modes

There are two join modes you can choose:

• concatenate: Simply combines the answers into a single list.
• merge: Aggregates scores of individual answers. With this mode, you can also specify the weights parameter, which assigns importance to the input nodes. By default, all nodes have equal weight.

📘

Answers from LLMs

Note that answers coming from a large language model (LLM) have no score. Set sort_by_score=False to join the LLM's answers with answers from other nodes.
For more information, see the Arguments section below.

Usage Example

One use case for JoinAnswers could be that your documents contain different types of data, for example tables and text. In your pipeline, you're using RouteDocuments to route the documents to different readers, depending on their content type. Each reader returns predicted answers, and you can then use JoinAnswers to produce a final single list of answers.

components:
  - name: Retriever
    type: TableTextRetriever
  - name: DocumentStore
    type: DeepsetCloudDocumentStore
    params: 
      embedding_dim: 768
      similarity: cosine 
  - name: Router
    type: RouteDocuments
  - name: TextReader
    type: TransformersReader
  - name: TableReader
    type: TableReader
  - name: JoinAnswers
    type: JoinAnswers
    params:
      join_mode: concatenate
      top_k_join: 10
      
      ...
      
 pipelines:
  - name: query
    nodes:
      - name: Retriever
        inputs: [Query]
      - name: Router
        inputs: [Retriever] 
      - name: TextReader
        inputs: [RouteDocuments.output_1]
      - name: TableReader
        inputs: [RouteDocuments.output_2]
      - name: JoinAnswers
        inputs: [TextReader, TableReader]
        ...

Parameters

You can specify the following parameters for FileTypeClassifier in the pipeline YAML: