📄️ Introduction
The data science team had made a number of end-to-end projects including the ones documented here. We needed a unified interface to interact with all of them since dashboards were cumbersome to both build and use.
📄️ Model
The model used for SynBot was GPT35-pss. It was made using Azure OpenAI and LangChain.
📄️ Prompt Engineering
A prompt template is made with LangChain for prompt engineering.
📄️ Cleaning
The first step is to break down the given prompt into a sequence of simpler prompts that are easier to process.
📄️ Type of prompt
The type of the prompt is detected next. This is an individual prompt in the sequence of instructions returned by the cleaning tool.
📄️ Retrieval Augmented Generation
The model for answering the prompts may require specific data to answer the queries. For example, if asked a question about forecasts, the model requires the actual forecast data.
📄️ Response history
The chat history must be passed to the cleaning agent for context. In case of a series of instructions rather than a single one, the responses to the intermediate tasks may be required as context for the subsequent tasks, so also must be passed as history.
📄️ Final Response
The response from the tool used to process the final instruction may not be human-friendly, only containing a single number or such.
📄️ Data Visualization
An llm does not have the innate capability to make dynamic visualizations like charts and KPI cards.
📄️ Interface
A custom GPT chatbot was made for each of the projects as an addon to the corresponding dashboard, including for Inventory Management, Document Retention and Crop Identification. From the master dashboard, the user can either go to the dashboard or the natural language interface.