microsoft/graphrag

Choose this fork if your priority is broader LLM-provider compatibility and more pragmatic example workflows, especially for Chinese ecosystem models. Choose upstream if you want the fastest-moving, better-maintained base with newer fixes and release cadence.

Choose this fork if you want a .NET-native GraphRAG variant and can tolerate losing some upstream Python workflow compatibility. Stick with upstream if you need the original Python CLI, notebooks, and fastest access to Microsoft’s latest fixes.

Prefer upstream unless you specifically need the fork's custom parsing, local search tuning, or visualization/demo additions. This fork looks useful as an experiment or reference implementation, but it is too stale and diverged for low-risk adoption.

Prefer upstream unless you specifically need this mirror. This fork does not show meaningful product additions, and it is already slightly behind upstream, so it offers little adoption advantage beyond being a close copy.

Choose this fork only if its Tongyi Qianwen/provider-specific changes are the reason you need it. For most adopters, upstream microsoft/graphrag is the better base because it is far more current, actively maintained, and likely to be easier to support.

Choose this fork only if its local test, workflow, or data customizations are specifically valuable to you. If you want current GraphRAG features, fixes, and documentation, upstream is the better default because this fork is materially stale and diverged.

Prefer this fork only if its specific tuning choices match your use case and you are comfortable owning an aging, highly divergent codebase. For most adopters, upstream GraphRAG is the safer choice because this fork is stale and materially behind on fixes and releases.

Prefer upstream unless you specifically need this fork's prompt-tuning/autotemplating changes or want to build on an older snapshot. For most adopters, the fork is too stale and too far behind current GraphRAG maintenance to be the safer default.

This fork looks like an inactive mirror rather than a differentiated downstream product. Prefer upstream unless you specifically need a frozen 2025-era snapshot to build from or reproduce against.