HeavyIQ leverages the power of a custom-trained Large Language Model (LLM) to allow you to ask questions of your data using natural language.
HeavyIQ is a new capability in the HEAVY.AI platform that allows users to use natural language to ask questions of and interact with their data. At the core of HeavyIQ is HeavyLM, a powerful Large Language Model (LLM) that has been custom fine-tuned on over 60,000 instruction pairs to deliver state-of-the-art performance for text-to-sql, natural language summarization, and other data analytics tasks such as sentiment analysis, named entity extraction, and classification. Since HeavyIQ is built around a custom model, it does not require online access to third-party LLM providers, ensuring complete data privacy and even allowing for fully air-gapped, offline deployment if desired.
In addition to the custom HeavyLM model, support for HeavyIQ is deeply integrated into both the HeavyDB database and Heavy Immerse visual analytics platform.
In HeavyDB, users can use the new LLM_TRANSFORM operator to call the HeavyLM model inline from SQL for a wide variety of use cases, whether ELT (Extract Transform Load) tasks like date value cleanup, or core analytics workflows like sentiment analysis or classification. The HeavyLM model is specifically trained to deliver terse responses suitable for downstream analytics, and the LLM_TRANSFORM operator allows optional constraint of the LLM output to a regex or a set of output values to ensure the results of the operator can be reliably used downstream.
From Heavy Immerse, users can work directly within the new SQL notebook to ask questions of their data in natural language, getting back SQL, and if desired, the raw or natural-language summarized results and compelling auto-generated visualizations. Unlike other text-to-SQL platforms, all relevant metadata, including top values, column ranges, table and column metadata, and with 8.1, user-provided guidance is automatically and intelligently fed into the prompts sent to the HeavyLM model to provide the model with the necessary context to deliver the most accurate answers.
How to Try it?
Learn more about HeavyIQ and try our interactive demos live here. HeavyIQ (using a publicly hosted model) is also available to all Free Edition users, and can be deployed privately behind your organization's firewall as part of an Enterprise deployment.
Guidance snippets are hints, tips, formulas or acronym explanations which you can optionally provide to the HeavyIQ language model. They are particularly useful in describing things which are not known to the base model but are important in your particular analytical domain. For example, the acronym TVD can stand for many different things, including “total variation distance” in statistics. But in the Oil and Gas industry, it stands for “true vertical depth.” Similarly, you may have industry or even company-specific KPIs or key performance indicators. These are all good candidates for inclusion as guidance snippets.
When HeavyIQ considers your analytical questions, it consults the guidance snippets as part of the query generation process. So in addition to proactively adding snippets based on domain knowledge up front, you can also add them “ad hoc” to clarify points which appear to be confusing the model. Administrators may wish to periodically review any unsuccessful queries and clarify terminology within them using snippets. This allows the model to learn relevant new information over time.
You can access the Guidance panel by clicking the “Guidance” tab in SQL Notebook. The Guidance panel displays a list of all guidance snippets associated with the current database, and allows superusers to create, edit, or delete snippets.
Guidance snippets can be sorted by oldest or newest updated date from the “Sort by” menu.
To create a new snippet, click “Add guidance snippets”.
This will give you an input to enter your snippet text. You can create up to 20 snippets at once by clicking the “+” button to add a text input for each additional snippet. Note that each snippet must be unique, and creation will fail if the content of a snippet is identical to any existing snippet.
The snippets list shows a truncated preview of each snippet, along with the time it was last updated. From the snippets list, click on an existing snippet to view its full text. From here, superusers can also modify the snippet.
To delete snippets, select the snippets you would like to delete by clicking the checkboxes next to them and click the “Delete” button that appears.
For a short period of time, you can restore the most recently deleted batch of snippets by clicking the “Undo” prompt displayed in the confirmation snackbar that appears.
