Fight AI Hallucination in Cesium Development — Your Feedback Is Needed!

General
1

Hi Cesium Community,

We’re currently exploring ways to enhance the development experience with Cesium by integrating AI assistants into the workflow.

As part of this effort, we’ve been experimenting with Context7 ( Context7 - Up-to-date documentation for LLMs and AI code editors ), an MCP (Model Context Protocol) server that provides AI tools with real‑time access to the latest documentation, source code references, and API details.

At the moment, Context7 has five Cesium‑related documentation sources and repositories indexed, which we believe can help significantly reduce AI hallucinations during Cesium development:

Documentation Original Context7 Context7 Library ID
CesiumJS API Reference CesiumJS API Context7 CesiumJS /websites/cesium_learn_cesiumjs
Cesium Repository GitHub cesium Context7 cesium /cesiumgs/cesium
Cesium for Unreal GitHub cesium-unreal Context7 cesium-unreal /cesiumgs/cesium-unreal
Cesium for Unity GitHub cesium-unity Context7 cesium-unity /cesiumgs/cesium-unity
3D Tiles Specification OGC 3D Tiles Context7 3D Tiles /websites/ogc_cs_22-025r4

We’ve opened a GitHub PR with details and a wiki on how to connect your AI assistant with the Context7 MCP server:
:backhand_index_pointing_right: https://github.com/CesiumGS/cesium/pull/13160

We’d love to hear your feedback:

  • Has Context7 (or MCP‑based tooling in general) been useful in your Cesium development workflow?

  • Are there other AI tools or approaches that you’ve found especially helpful?

  • Have you encountered situations where Context7 (or similar AI tools in general) doesn’t provide the expected support?

    • Specific examples or test cases would be extremely helpful for future improvements.

  • Are there any concerns or considerations we should keep in mind as we continue exploring this?

Your feedback would be greatly appreciated!

1 Like
Cesium Visual Studio Code Extension: We need your input!
FEEDBACK REQUESTED: Native AI Copilot for Sandcastle
2

Thanks for doing this! I’m going to start playing with the Context7 MCP as I get time.

My one thought, having not used Context7 yet, but also having done quite a bit of LLM development with CesiumJS is that I’m not sure Context7 is needed. Early on, (models prior to 4o), ChatGPT had some confusion with the CesiumJS class structure. That quickly faded away though. I’ve built several apps with very little to no LLM hallucuination or confusion.

Would it be useful for me to create a repo of the transcripts for those development sessions for reference, or provide some other sort of documentation?

Best Regards,

Hamilton Carter

415-799-1629

3

Thanks—that’s great to hear, and your experience matches what I’ve seen. The newer models handle CesiumJS much more reliably.

Where the Context7 MCP server can still help is when new APIs, features, or examples are added to Cesium. It gives the model immediate access to the latest changes instead of waiting for future training cycles.

Any format of examples or transcripts—would be really helpful. Seeing real interactions gives us a clear picture of where the base model is already strong, where Context7 MCP actually helps, and where (if at all) it still fails. That would also tell us whether we need to build a more customized Cesium‑focused MCP layer to cover any remaining gaps.