nmachado

We actually leverage LLMs very sparingly. We do not generate the transformed data directly as this would not just introduce significant latency, but also quality and reliability issues. Rather, we use LLMs to produce high-level mapping/transformation logic in a language of our design that is deterministically executed to produce your desired data. So this means LLMs are used only when you introduce new data formats that requires new logic, and used surgically even then. The vast majority of usage so far in terms of volume leverages the logic created in the underlying pipeline and will not have latency issues. This also allows for building reliable and stable pipelines with our APIs, which is a requirement that's difficult to meet considering the non-determinism of LLMs.

We haven't had bandwidth to run it on this example yet, but will report back when we do!

Sounds very useful, thank you for sharing.

Customers can use Lume to do one-off mappings along with recurring mappings. As seen in the demo video, you can create a mapping between any source and target schema. So, the workflow of using Lume is the same for both one-off and recurring cases. Was this what you were referring to?

Thank you!

wow! It's a pleasure, qarl :) Reach out anytime.

They are written in python but SQL is in the roadmap.

Great question. We focus on embedding in your data pipelines themselves. So, our AI automatically maps data, and can be used as a data pipeline indefinitely. Indeed, it can connect to APIs and handle dynamic output or edge cases you did not expect. Also, we work on handling any complexity of transformations (1-1 mappings, all the way to string manipulation, classification, aggregrations, etc).

Thank you for shouting this out! I'll look into getting a smaller version in there.

Right Lume is a low to no code tool (see demo), but we have gotten requests for an SDK. Creating a library / SDK is in our radar! If anyone has personal or company use cases for a library / SDK, please email [email protected], we'd love to learn more.

Gotcha! In that case, we do not yet support an end-to-end experience for this, but would be willing to prioritize building it for clients if we get strong demand.

This would be a concern if we are feeding the raw user input and feed it directly into an LLM. In our case, we are not simply a wrapper over an LLM.

There are multiple parsing and rule-based steps done to the input schemas - we extract specific pieces from the schemas and convert them to our internal format before feeding it our models. Thus, it mitigates such malicious behavior.