LlamaIndex

We could build this application using ordinary code and the APIs of the systems involved.

If the retrieval and the data side becomes particularly sophisticated, we might use something like Lama Index to help with that layer.

If the application has a stateful workflow with several branches, loops, and checkpoints, something like Landgraf may help us represent and manage that execution.

We might use a provider SDK to manage model calls, tools, sessions, or tracing.

I love information retrieval and this kind of setup, but I think also generalizing these ideas to like a text to SQL type of setting and thinking about structuring.

Like there's this new, like Llama Index has been doing this work on Llama Parse and they have this thing where you give it some PDFs and then it will predict a schema, like a structured schema, structured relational schema.

And yeah, I guess I think that one's really interesting because I think there's just like a lot you can do with these kind of structured schemas, especially for like at large scale, of course, like large scale symbolic aggregations or even, I think like the cool thing about like data lake houses and stuff is when you're separating the data and you have maybe like different storage tiers for different kinds of data.

Like, oh, this one's got like hot storage.

None of it is new.

It has shipped as a named retriever in the LangChain and LlamaIndex documentation since twenty twenty-three.

A raw table embeds poorly, while a plain language summary of it embeds well.

So you have a model read the table and write a short description of it like Q3 regulatory capital disclosure showing a Tier one capital ratio of twelve.

After that, you orchestrate.

Tools like LangChain or LlamaIndex or a custom microservice layer define the order and logic of how your agents interact.

And critically, the final output surfaces inside the research portals and dashboards your attorneys already use.

If the workflow is friction-heavy, adoption dies.