ROADMAP.md

stanford-futuredata/ColBERT Roadmap

Guiding Philosophy

stanford-futuredata/ColBERT should remain a stable, canonical reference implementation of late interaction, especially for newcomers to late interaction and the GPU-poor*. Folks looking for the bleeding-edge of late interaction should check out the fantastic lightonai/PyLate library.

*stanford-futuredata/ColBERT's coupled model-index design enables batched encoding with immediate compression, maintaining sub-5GB memory usage even for multi-million document collections.

Immediate Goal: Dependency Stabilization (~3 months)

• Upgrade PyTorch to 2.x
• Upgrade transformers (remove deprecated AdamW)
• Replace faiss with fastkmeans
• Test Python 3.9-3.12 compatibility
• Resolve crypt.h/ninja errors.
• Merge distributed training fix in #258 (potentially related: #132 and #233)
• Replace git-python with GitPython in PyPI (already changed in repo in commit 736f88b)

Medium-Term Goals: Documentation and Bug Fixes (~6 months)

• Update documentation
  • Address documentation updates in issues/PRs (#316, #153, #167, etc.).
  • Create an llms.txt and llms_ctx.txt for the repo.
• Investigate issues:
  • Bug (#159, #317, #360, etc.)
  • Training (#262, #265, #291, etc.).
  • IndexUpdater (#180, #261, #276, etc.).
  • Multi-GPU (#158, #265, #318, etc.).
  • Ready-to-close issues after review/repro (#139, #179, #335, etc.).
  • etc.

Long-Term Goals: Feature Requests (~3 months)

• Resuming training from checkpoint (#307).
• Allow string pids (#326).
• Explore batch size handling options to resolve OOM during search.
• etc.