-
Notifications
You must be signed in to change notification settings - Fork 25
Improve AI responses on docs #344
Copy link
Copy link
Open
Labels
effort: highLarge or difficult task.Large or difficult task.priority: 2We will do our best to deal with this.We will do our best to deal with this.type: perfChange that improves performance.Change that improves performance.work: complicatedSense-analyze-respond. The relationship between cause and effect requires analysis or expertise.Sense-analyze-respond. The relationship between cause and effect requires analysis or expertise.
Description
Metadata
Metadata
Assignees
Labels
effort: highLarge or difficult task.Large or difficult task.priority: 2We will do our best to deal with this.We will do our best to deal with this.type: perfChange that improves performance.Change that improves performance.work: complicatedSense-analyze-respond. The relationship between cause and effect requires analysis or expertise.Sense-analyze-respond. The relationship between cause and effect requires analysis or expertise.
Custom Prompt for Algolia
Play around with the Algolia system prompt to identify the best instruction for creating good responses in the docs.
A best instruction, imo, should create responses as insightful and helpful as us (the team). A strong response should go to lengths beyond answering the direct question, it should also provide additional indirectly helpful information to the user.
For example:
Q: Can I create a stream with a non-transferable token?
An excellent answer would mention that its possible, but also include additional details such as importance of whitelisting the contract, specify contracts for both Lockup and Flow, highlight potential issues such as user unable withdraw funds if transferability is disabled etc.
This Prompting 101 from Anthropic could be a useful.
Selection of OpenAI Model
Experiment with different models from OpenAI. You can select them in the Algolia dashboard.
The pricing for different models can be reviewed here. The objective is to identify the most cost-effective model that still delivers high quality responses. Some of OpenAI's latest models, such as GPT-5 mini, are even more affordable and better than the GPT-4o model (which is currently used).