DiaKoP: Dialogue-based Knowledge-oriented Programming for Neural-symbolic Knowledge Base Question Answering (CIKM 2024, Demo Track)

Introduction

We proposed DiaKoP, a multi-turn KBQA system that consists of two main modules. The Dialogue-based Explainable Interactive Frontend features a chat interface that allows users to interact with a predefined knowledge base (a) and a program editing interface that enables users to correct incorrectly parsed logic forms (b). The Multi-turn Question Answering Backend is the core component responsible for processing and managing the conversation. It comprises several sub-modules: the Dialogue History Tracker (c) and Dialogue Policy (d) manage the conversation flow, while the Knowledge Source (e) stores the factual information to answer user's question.

Prerequisite

Model

Download the following model from huggingface

1. Quantized Llama 3 - 70B Model Hugginface link
   After downloading, modify env MODEL_PATH in launch_worker_vllm.sh

2. Semantic parsing model Huggingface link
   After downloading, modify env CKPT_PATH in launch_semantic_parser.sh

Environment Setup

# create conda environment
conda create --name diakop python=3.10

conda activate diakop 
pip install -r requirements.txt
conda install graphviz

Run

Activate conda and change directory to the scripts_bash

conda activate diakop
cd scripts_bash

RUN SCRIPTS BELOW IN DIFFERENT SHELL

1. Launch the controller.
This controller manages the distributed workers.

sh launch_controller.sh

2. Launch the model workers.
Before running, modify MODEL_PATH and NUM_GPUS, GPU_IDS variables in launch_worker_vllm.sh. Wait until the process finishes loading the model and you see "Uvicorn running on ...". The model worker will register itself to the controller.

sh launch_worker_vllm.sh

3. Launch semantic parser.
Befor running, modify CKPT_PATH and GPU_ID variables in launch_semantic_parser.sh.

sh launch_semantic_parser.sh

4. Launch the Gradio web server.

sh launch_gradio.sh

5. Launch Program Editing Interface.
Make sure you have installed required packages before running code below.
Refer README for package installation guide.

cd DiaKoP/frontend-src/KoPL-vis/vis
npm run serve

6. Deploying KoPL execution engine.
Please refer Github on how to deploy a KoPL program execution engine. After deploying the engine, modify the API link in the env variable KOPL_EXEC_ENGINE_API in launch_gradio.sh.

Note

To use glm-3-turbo api, create .env file and add your api key, example below:

CHATGLM_PRO_API_KEY=<YOUR API KEY>

This API does not adopt the proposed multi-turn question-answering backend.