Add support for TiDB Cloud (TiDB Cloud / TiDB Zero) as a database backend

[you06]

Hi maintainers — thanks for the project!

I’d like to ask whether you’d be open to adding TiDB Cloud as a supported database backend, and whether contributions in this direction would be welcome. In particular, TiDB Cloud offers a serverless option that seems very friendly for users who want to try the project quickly.

• No sign-up required to start a database with a 30-day lifetime (TiDB Zero): https://zero.tidbcloud.com/
• After sign-up, the same database can be kept indefinitely for free (free tier / Starter).
• Built-in Text Embedding capabilities, which reduce CPU requirements for deployment.

I’d be happy to contribute this feature to the project. If you think it’s feasible, I can propose a design and discuss with you before start implementation.

[kwaa]

Contributions are welcome, but that (TiDB support) will likely be sometime after version 0.1 or even longer.

It seems TiDB is MySQL-compatible, so compatibility might not be too difficult, but we're currently evaluating adding Graph feature. (this will cause a breaking change)

[Lilia-Chen]

Yeah... a very huge decision is still TO-BE-MADE about how we are going to support graph related functionality, whether using stand alone graph engine or more compact solutions...

[you06]

Thanks for the notice of graph feature. Are you planning to orchestrate the memory contents in a graph so that it can easily search the close memory sections.

I’ll wait for your decision, take your time.

[Lilia-Chen]

In my view, @kwaa may have different ones hhh, graph is a natural way to perform the search of memory, but how exactlly we organise them is still quite comlicated...

• Are we going to pre-defined the layered structure? or treat the atom as a node and then dynamically link then based on the needs?
• What's the node of the graph, we do have segmentations to split the raw message streams, but after that how are we going to abstract this information? event by event? or by entity (what kind of entities? name, place, time, etc?)
• Do we need a graph engine or just traditional SQL table and mimic the graph structure? (this will impact what db we can support)
• Graph search requires recursitive process, what's the cost and benefits? especially when we need to search quickly (we can write slowly, which is fine)

tbh I haven't got a very systematic idea about this yet... my current focus is on some detailed designs that can boost benchmark performance. I realised that some differences in details (no change in ideas or general design) can still impact a lot on the final performance.

[siddontang]

Hi

The head of TiDB here, from my side, for the graph feature, based on previous experiences, and I guess your data for each user will not be too much, I think using a table with mimic graph structure like we do in tidb.ai is enough, you don’t need to maintain another independent graph database.

If you want, we can have a deep discussion further.

Another value I think we can provide is that the default TiDB cloud free tier can provide 5GB storage and it may meet most of the users’ needs. Even the users save more than that, TiDB cloud can still guarantee a very low cost.

Btw, we also believe in the future everyone must have some of their own AI friends, so thanks for the project.

[YangKeao]

After reviewing the codes, I think it'd be better to firstly have an abstraction storage layer before importing any new storage solution (or the code will turn into pure spaghetti). I've submitted a proposal #43 PTAL.

The corresponding codes will be submitted later.

Graph search requires recursitive process, what's the cost and benefits? especially when we need to search quickly (we can write slowly, which is fine)

@Lilia-Chen BTW, I had some magic thoughts about implementing graph search with recursive cte 🧙‍♂️ , as a SQL mage.

[Lilia-Chen]

Hi @YangKeao @siddontang @you06! Thank you very much for being interested in our project.
We have a Discord channel for plast-mem dev, and you are welcome to join and have further discussion ~
https://discord.gg/FkDxSk26Z4