Skip to main content

LLMs

Adding LLMs to our dataflow​

Let's add additional operators:

wget https://raw.githubusercontent.com/dora-rs/dora/v0.3.2/examples/python-operator-dataflow/keyboard_op.py
wget https://raw.githubusercontent.com/dora-rs/dora/v0.3.2/examples/python-operator-dataflow/microphone_op.py
wget https://raw.githubusercontent.com/dora-rs/dora/v0.3.2/examples/python-operator-dataflow/whisper_op.py
wget https://raw.githubusercontent.com/dora-rs/dora/v0.3.2/examples/python-operator-dataflow/sentence_transformers_op.py
wget https://raw.githubusercontent.com/dora-rs/dora/v0.3.2/examples/python-operator-dataflow/llm_op.py
wget https://raw.githubusercontent.com/dora-rs/dora/v0.3.2/examples/python-operator-dataflow/file_saver_op.py

and the dataflow configuration:

nodes:
  - id: webcam
    operator:
      python: webcam.py
      inputs:
        tick: dora/timer/millis/50
      outputs:
        - image

  - id: object_detection
    operator:
      python: object_detection.py
      inputs:
        image: webcam/image
      outputs:
        - bbox

  - id: plot
    operator:
      python: plot.py
      inputs:
        image: webcam/image
        bbox: object_detection/bbox
        line: llm/line
        keyboard_buffer: keyboard/buffer
        user_message: keyboard/submitted
        assistant_message: llm/assistant_message

  ## Speech to text
  - id: keyboard
    custom:
      source: keyboard_op.py
      outputs:
        - buffer
        - submitted
        - record
        - ask
        - send
        - change
      inputs:
        recording: whisper/text

  - id: microphone
    operator:
      python: microphone_op.py
      inputs:
        record: keyboard/record
      outputs:
        - audio

  - id: whisper
    operator:
      python: whisper_op.py
      inputs:
        audio: microphone/audio
      outputs:
        - text

  ## Code Modifier
  - id: vectordb
    operator:
      python: sentence_transformers_op.py
      inputs:
        query: keyboard/change
        saved_file: file_saver/saved_file
      outputs:
        - raw_file

  - id: llm
    operator:
      python: llm_op.py
      inputs:
        code_modifier: vectordb/raw_file
        assistant: keyboard/ask
        message_sender: keyboard/send
      outputs:
        - modified_file
        - line
        - assistant_message

  - id: file_saver
    operator:
      python: file_saver_op.py
      inputs:
        file: llm/modified_file
      outputs:
        - saved_file

Try it out with:

dora up
dora start dataflow.yml --attach

record speech to text​

The keyboard will record every key input of the keyboard.

If the keyboard receive submit the keyword record, the microphone is going to be triggered and the audio is going to be transcripted using OpenAI whisper.

change source node​

The code modification flow works by first submitting an instruction with the keyword change. The instruction is then going to be passed on the vectorDB to retrieve the closest node source code for it to be changed.

The source code and the instruction is then fed to an LLM to be modified, which is then saved into the source code.

This later triggers a hot-reloading sequence for the node to be reloaded.

The end result should correspond to the instruction.

ask a question to the assistant​

You can ask a question to the assistant by submitting the keyword ask.

The question is then going to be passed to the LLM which will reply in the window.

send a message to the dataflow​

You can also send a message directly to the dataflow by using the keyword send and then specifying the data that you want set.

Note that the data and topic should be specified both in the dataflow and in the LLM node to send data`

Getting started​

You can try the following instruction:

ask how are you
change bounding box plot to red
change confidence value to percentage
change object detection to only detect person
send 200 200 200 400 to topic line
record # Then speak