Below is a very high-level roadmap for HeimdaLLM.
Want to know when these features get implemented? Add your email to this form and I will contact you.
More SQL statement types
SELECT gets the biggest bang for the buck, but other SQL statements are also very
useful and would benefit tremendously from a Bifrost. For example,
traverse("Add a new calendar entry for Dinner at 7 on friday")
Could produce a validated SQL query:
INSERT INTO calendar (title, when, user_id) VALUES ('Dinner at 7 on friday', '2023-07-01 19:00:00', 123)
Generalized constraint spec
The current implementation requires a Python application to use HeimdaLLM, because constraint validators are defined by subclassing a Python class. A future implementation could be language agnostic by providing an api and a JSON or YAML spec for constraining LLM output.
More LLM integrations
Currently we support OpenAI, but I intend to add support for all major LLM API services, and private LLMs, as they become more capable.
I will be adding support for more SQL-based databases:
Want us to prioritize a specific database? Let us know by voting here.
Bifrosts are not limited to converting human input to trusted SQL statements. HeimdaLLM is generalized enough to support many kinds of structured output. I intend to develop more Bifrosts that facilitate natural language interactions with your application. Stay tuned!