Luzmo IQ is designed to work with any data source, data structure, and within any industry or context. While this is a great starting point, there are several options to influence how Luzmo IQ answers questions. Tweaking your Luzmo IQ setup to match the expectations of your customers will greatly improve its effectiveness.
Luzmo IQ Vector Embeddings will automatically create and store vector embeddings for your datasets. Vector embeddings are numerical representations of hierarchical data, capturing their meaning and relationships. They help Luzmo IQ understand context and similarity, enabling efficient search and semantic analysis.
Enable vector embeddings by navigating to the details page of your datasets and click "Activate" at the Luzmo IQ Vector Embeddings section.
Vector embeddings will improve search within hierarchical columns. For example, searching for a name like "Emma Johnson" might fail without vector embeddings if the name is formatted or written slightly different in the dataset, like "Emma Grace Johnson" or "Johnson, Emma". With vector embeddings enabled, Luzmo IQ will find the correct values through semantic search and, if uncertain, confirm the result: "Did you mean Emma Grace Johnson?".
Vector embeddings enable more complex queries. For example: Assuming a dataset with employees and their country will allow questions like "What's the total amount of sick leave days of Europe-based employees", where "Europe-based" is a selection that Luzmo IQ will perform with help of vector embeddings.
When you find Luzmo IQ struggles to select relevant data from hierarchical columns, check if you've enabled the feature "Luzmo IQ Vector Embeddings" for the relevant datasets.
The first step in Luzmo IQ's thought process is to select datasets and columns that can be used to answer a question. This selection is performed based on the dataset names and column names. Using descriptive column names that prevent confusion with other columns will help a great deal in selecting the correct columns to answer a question.
Examples of descriptive column names:
order_quantity instead of count
order_date instead of data
customer_full_name instead of name
product_category instead of cat
Note that the format of the dataset or column name is not important. Any common naming format, like snake case or camel case, will work with Luzmo IQ.
A useful rule of thumb: When a (smart) human is able to explain the difference between your dataset columns just by looking at their names, Luzmo IQ will be able to do the same.
Luzmo IQ takes into account the data types of your columns when selecting columns or formatting data in charts and text. It's important to set these types correctly. For example: When asking questions about a certain timeframe, Luzmo IQ will look for a date or datetime columns. Dates stored as hierarchy type can't be used to answer such questions.
Setting correct subtypes is also important. When a column contains currency values, make sure the column's type is set as currency rather than just numeric. The same applies to numeric data that represents duration.
Luzmo IQ is able to create formulas to answer questions. For example, Luzmo IQ will easily answer questions about profit margin when a dataset contains only revenue and cost information. However, many metrics have different methods to be calculated, resulting in a different outcome. For profit margin, one could calculate gross profit margin, net profit margin, operating profit margin, etc..
To improve the success rate and consistency of formulas, you can add your own formulas to a dataset. Luzmo IQ will automatically prioritize existing formulas over creating its own.
It's wise to add a formula for each commonly used metric that's not present in one of the dataset columns. Make sure to give them a descriptive name. When your customers are used to using abbreviations, it's most effective to add the full and abbreviated name. Example: Net Profit Margin (NPM)
Luzmo IQ allows you to programmatically add context to the LLM prompt through the embed token. This gives you tremendous control in steering Luzmo IQ towards specific outcomes, without exposing these instructions to your users. The following tips will help to write an effective prompt:
Example 1: Adding business context
Example 2: Add context to help to select columns for common questions.
Assume a "Projects" dataset with has several columns with cost information, like actualCost, budgetedCost, or billedCost, you can prevent Luzmo IQ to make unexpected selections.
Option 1: Prescribe a default
Option 2: Force follow-up questions
Example 3: Add instructions on how to create text and chart answers
Set default periods for displaying time-based data
Provide instructions for writing large numbers
You can personalize the custom prompt for individual users through the embed token. In the above example, this allows you to instruct the use of long-scale or short-scale number notation depending on localization settings.
Luzmo IQ is set up to generate answers as fast as possible without sacrificing accuracy and consistency. The speed with which answer are generated depends a great deal on the size of the datasets connected. There are a few options to positively influence answer speed:
In most cases, Luzmo IQ serves a different purpose than dashboards. Often, users want to understand events that are happening now. With that in mind, it might not be necessary to feed the full data history of your datasets to Luzmo IQ. Consider setting an embed filter on the embed token to limit your data to, for example, the past year.
Alternatively, consider adding a custom prompt that instructs Luzmo IQ to only consider a recent period by default. This will still allow users to ask questions outside that default period.
A data source that's not optimized for analytical queries will have a bigger negative impact on Luzmo IQ than other Luzmo charts. To formulate a single answer, Luzmo IQ might do multiple queries.
Consider using Warp to speed up queries made by Luzmo IQ.