Customizing Prompts

Implementers of the trait SimplePrompt take a Prompt as argument. A Prompt references a PromptTemplate and can render, with optional context, to something that goes to an LLM.

Avoid creating compiled templates during a pipeline or under load. The repository is managed internally by a RwLock and this will negatively impact performance. Use ad-hoc templates from strings instead.

Customizing templates

Templates use tera under the hood and can be fully customized.

For example:

let template = PromptTemplate::try_compiled_from_str("hello {{world}}").await?;
let transformer = MetadataQACode::builder().client(...).prompt_template(template).build()?;

When querying, templates are converted to prompts, have their context added, and are then send to the client.

By default, Swiftide transformers have the full Node available as context, and any other values the transformer provides.

Ad-hoc templates

Prompts can also be constructed ad-hoc from a &str or String, with optional context.

let prompt: Prompt = "hello {{world}}".into();

openai.prompt(prompt.with_context_value("world", "swiftide")).await?;

Bring your own

The prompt repository can also be extended with your own tera instance. This can be useful if you have your own templates in markdown, or other functionality you would like to add.

You can also overwrite all the default prompts this way, by overwriting the template names.

See the api documentation for more details and examples.