Summary Prompt engineering davinci-003 on our own docs for automated support (Part I) | Patterns www.patterns.app
1,538 words - html page - View html page
One Line
Patterns Data Systems employed an LLM support bot using OpenAI's GPT-3 davinci-003 engine and Slack to respond to tech support questions with knowledge from text corpus, Slack channels, and previous support tickets.
Key Points
- Patterns employed an LLM support bot using OpenAI's GPT-3 davinci-003 engine to scale technical support.
- Prompts were created from the text corpus to generate labelled training data for the model.
- Python nodes were used to process webhooks and transform them into desired outputs, which were then sent to external systems.
- Data was prepped and uploaded to OpenAI, then fine-tuned to configure a Slack bot.
- Part II of the post will explore ways to improve the bot's accuracy, such as expanding its corpus and using embeddings.
- Users can create a free Patterns account bot and a shared Slack channel to interact with it.
Summary
260 word summary
Prompt engineering davinci-003 on our own docs for automated support (Part I). To scale technical support, Patterns employed an LLM support bot using OpenAI's GPT-3 davinci-003 engine. To generate labelled training data for the model, a series of prompts were created from the text corpus. The result was a Slack bot that can respond to tech support questions with knowledge from the text corpus, Slack channels, and previous support tickets. Feed GPT a chunk of data and ask it to generate three relevant questions. Feed these questions back to GPT and ask it to answer them using the chunk, creating labeled training examples. Import training data into Patterns using Airtable components. Generate questions as prompts from docs using OpenAI Completions component and generate answers for these questions. Python nodes are used to process real-time webhooks and transform them into desired outputs, which can then be sent to external systems. Data was prepped and formatted for OpenAI's CLI data preparation tool, then uploaded and fine-tuned. A Slack bot was configured using the fine-tuned model, resulting in relevant answers. In Part II of this post, we will explore ways to improve the bot's accuracy, such as expanding its corpus and using embeddings. If you have questions, contact us at [email protected] or create a Slack channel from your dashboard. Check out the code for a free Patterns account bot and create a shared Slack channel to interact with it. Prerequisites: generate training data, upload to OpenAI, configure model as a Slack Bot. Learnings and next steps. Copyright © 2022 Patterns Data Systems, Inc.