[rogerkirkness]

Convictional founder here. Our experience is different than others:

- We had to sync, pre process and index the data to make the resultant knowledge search outputs actually good. MCP totally fails at this by comparison.

- It is not hugely painful thanks to bulk APIs, in Gmail in particular, as well as webhooks. We implemented both of them and it works well (so far).

- We wired it all up ourselves. Given the conclusion we had about pre-processing and indexing being required to make it work well, this seems preferred.

I think that MCP and using an integration platform will ultimately not work for any kind of agentic or deep research task heavily depending on Gmail context.

[Bahushruth]

That’s really interesting, especially your point about preprocessing and indexing being required to make search outputs good. What was the first sign that made you realize querying live APIs wasn’t enough?

Was it latency, missing data, or just that results weren’t relevant? And when you say preprocessing, what kind of transformations or normalization ended up being most important?

[rogerkirkness]

Keywords or vector search on their own don't get good results for high entropy queries. MCP type approach is good for low entropy things like fact-based single source answers. [1]

[1]: https://arxiv.org/abs/2504.07106

[yashgupta417]

what is your company doing exactly?

[rogerkirkness]

Collaborative email, meeting recording, knowledge search and goal tracking in one thing. The search applies across emails and meetings, but also other things. We had to figure out whether it would be sufficiently good to connect third party tools, and basically concluded no. We did some research to understand why (1).

[1]: https://arxiv.org/abs/2504.07106