Develop a semantic layer

[wiicode]

I would love to see a configurable semantic layer to define the purpose of some of my data structures. I built complex AirTables with many relationships, many agents, and thousands of rows of data. I have been attempting to find ways to get high-level summaries and keep running into limits. While I will continue to seek out ways to get more insight out of my data, I feel an MCP semantic layer that I could manage would be a band-aid until better options come to exist.

[domdomegg]

Hmm that sounds interesting - are you able expand more on what you think this semantic layer would look like?

[wiicode]

I don't know just yet. For now I am using schema files and explanations. The Claude Skills may change how I do this. My big issues have come from trying to ask bit questions to agents not in Claude Desktop. For example, AWS Bedrock (btw, reasoning fail after hours of development), MAKE, etc. That's where I think semantic layer makes more sense because those things are far from good for reasoning and deep insights. My current consensus is that all of it blows up context windows.

I am currently most likely able to call out the issues rather than solutions. I am working on the latter. Here are a handful of things from this week.... I don't yet know how to frame this as an issue, or perhaps I lack the skills to do so.

Imagine a database with many tables and many relationships between tables. AirTable AI is very good at understanding each row. AirTable AI is reasonably ok at understanding Linked fields (this is rapidly changing).

Pretend we have an operations database and I want to know if a customer side-loaded a Google account. DB stores population of people and a synch with google, the comparison is done, everything is good. This works fine with minimal direction.

Suppose you just want to know if a customer side-loaded a user.... without specifics.... things start to go off the rails. Suppose you want an agent to work that out daily....so not Claude Desktop.. things go even more off the rails.

Suppose you now wanted a list of side-loaded users daily across all customers... and in a db that is changing structure and being adapted with AI agent fields and other things... We're in this high-velocity environment where AI is changing.... and yet agents seem to really be stuck in this "precisely defined workflow mode" that is not helpful.

Really I want agents to be able to use information to create briefs and summaries. I want to ask an agent to do a complicated thing, let it run in chunks and give me a bunch of complex summaries. I can do this with iteration in Claude Desktop but that's seemingly my boundary... the moment I hand it over to automation I am back to sq 1. My hope is that being able to tell AI how to navigate the base I might be able to keep context windows from blowing up.

[NickCrews]

It sounds like the basic schema of "this table has a 1:N relationship with this other table" (which you could programmatically infer from the base schema) is inadequate. You really also need to store the business logic, eg "when a user doesn't have a billing account linked, that means they should not be able to access these tables". This sounds to me like you, as a human, are going to be responsible for writing and then updating those docs. I think any implementation should prioritize that experience: if it is difficult to keep those docs up to date, then you aren't gonna do it.

My first thought would be to store all this info in the into the descriptions of tables and columns. Then it is very easy for you to edit, and it is very easily programmatically accessible.

Would that work for you?

I think you are going to need to come up with a few more concrete examples. Those ones you gave aren't in enough detail for me to understand. What is a "side-loaded user"??

Also, there are two planes where this extra info could be used:

1. just as "suggestions" to the ai as prompts, but not strict enforcement of policies. You rely on the AI to be sane.
2. actual runtime enforcement of policies, so if the AI makes an illegal MCP call, it actually gives back useful info.

1 is going to be much easier I imagine, but less powerful.