Query Best Practices

Prompt engineering in Halcyon

To be able to write the most productive queries in Halcyon, you first need to understand how the system works.

When you run a query, Halcyon follows a two-step process to provide a well-researched response. First, it searches your filtered documents for the most relevant information — including text, images, tables and charts — to use in response to your question.

Then, AI extracts that relevant information and uses it to generate a query response, providing citations back to the original source material.

The majority of this section is going to focus on the second part - how the AI responds to your query, but we want to emphasize how important the first part is in the query/response process: taking the time to filter your searches carefully will narrow the universe of information such that the AI has the most relevant information with which to respond.

In other words: if you are getting bad/non-results, first refine your search criteria before refining your query/prompt wording.

For example, if you care about examples of natural gas power plants in Maryland, instead of using “Maryland” as a keyword, filter to the Maryland PUC in the Publisher drop down. Here you can see a side-by-side of the results that populate when filtered to Maryland as publisher vs Maryland as a keyword.

How to write better prompts

Writing effective prompts is a skill that is thankfully easy to master. Some questions to consider when writing your prompts:

Let’s go into detail about each of these guidelines:

Query context

Imagine you’re asking a brand new intern to do some research for you. What context would you need to include for them to be successful? Consider that the Halcyon app is like an intern that doesn’t know what you don’t tell it. Make sure to include all relevant details in your query.

Instead of this:

Can utilities own batteries?

Try this:

Do California investor-owned utilities have programs that allow them to own batteries at customers' homes in wildfire danger zones?

Totality queries

The system will find many examples of information related to your question, but it currently doesn’t execute an exhaustive search for all instances. For example, asking questions about all transmission lines will not yield a comprehensive study of all transmission lines. Asking for examples of notable transmission development typically yields more useful information that can guide further research.

Instead of this:

Find me regulatory filings associated with solar projects in California that have not been built.

Try this:

Find me some examples of regulatory filings associated with solar projects in California that have not been built.

Simplify, simplify, simplify

Once you’ve included all the necessary context, and avoided asking totality queries — now your best bet is to simplify. Breaking a research question apart into its component queries is a good way to narrow your query and get more meaningful results.

Instead of this:

I'd like you to find other parties that support or object to anything I may have said in my positions. Also, separately, please compile similar comments that parties have made and categorize them by topic.

Try this:

Summarize this comment and state whether or not it supports allowing data centers to "bring their own capacity."

Other tips and tricks