GitHub user yihua added a comment to the discussion: [DISCUSS] New Agentic 
lakehouse module & built-in AI agent gateway

+1, super excited about this direction.

A few things I'm curious about:

- Pluggable LLMs: how swappable is the model layer? Ollama out of the box is 
great; we could also add a clean interface to point the gateway at different 
local models (or vLLM/hosted endpoints) or remote onse without touching the 
rest of the stack. Is that the intent with the values-file config?

- Hudi as a storage layer for agents: beyond talking to tables, is it in scope 
for Hudi to be the backing store for agents built on frameworks like LangChain? 
A few directions I'd love to see explored:

(1) Retrieval: Hudi as a vector store / retriever using the vector search 
index, so agents pull context directly from Hudi tables at inference time.
(2) Unstructured data: Hudi can keep docs/embeddings and structured data side 
by side, indexed and queryable through the same table.
(3) Agent memory / state: Hudi as a durable store for conversation state and 
long-term memory, with incremental queries keeping that memory fresh and 
time-travel giving reproducibility for a given session or run.
(4) Traces + eval loop: writing agent traces back to Hudi closes the loop for 
eval, fine-tuning, and layout/index optimization and the table gets smarter the 
more the agent runs against it.

- Trino + caching: it's worth leaning on Trino's caching for the read-only 
query path; it should meaningfully cut latency on repeated agent tool calls and 
keep responses snappy.

GitHub link: 
https://github.com/apache/hudi/discussions/19264#discussioncomment-17644453

----
This is an automatically sent email for [email protected].
To unsubscribe, please send an email to: [email protected]

Reply via email to