flow-query.yml

jtype: Flow                                       # This file defines the query flow which is used for searching in the indexed documents
version: '1'                                      # The query flow is very similar to the index flow - only the differences are explained here
with:
  prefetch: 10
  port_expose: $JINA_PORT
  workspace: $JINA_WORKSPACE
pods:
  - name: loader                                  # Again, we start two paths in the flow - here we start the image path
    uses: ImageReader
    py_modules: 'flows/executors.py'
    read_only: true
    needs: [gateway]
  - name: image_encoder                           # Now, encode the images and compute the embeddings
    uses: 'jinahub+docker://CLIPImageEncoder/v0.1'
    volumes: $HOME/.cache/huggingface:/root/.cache/huggingface
    timeout_ready: 600000
    read_only: true
    needs: loader
  - name: text_indexer                            # Now, we use the text indexer in the image path - This is how we achieve the cross-modality here
    uses: 'jinahub://SimpleIndexer/old'           # The text indexer has indexed all text documents and stored them on disk.
    uses_with:                                    # Then we return the closest matches as results
     index_file_name: 'text_index'
    needs: image_encoder
    force: True
    read_only: true
  - name: text_filter                             # Here, the text path starts
    uses: TextFilter
    py_modules: 'flows/executors.py'
    needs: [gateway]
  - name: text_encoder                            # Compute the embedding of the search text
    uses: 'jinahub+docker://CLIPTextEncoder/v0.1'
    volumes: $HOME/.cache/huggingface:/root/.cache/huggingface
    timeout_ready: 600000
    read_only: true
    needs: text_filter
  - name: image_indexer                           # Now, we use the image indexer in the text path - this is again how we get cross-modality
    uses: 'jinahub://SimpleIndexer/old'           # The image indexer has indexed all images and their embeddings
    uses_with:
      index_file_name: 'image_index'
    force: True
    read_only: true
    needs: text_encoder
  - name: join_all                                # Wait for both paths to finish and join the results
    needs: [image_indexer, text_indexer]