Skip to content

Latest commit

 

History

History
39 lines (38 loc) · 2.24 KB

TODO.md

File metadata and controls

39 lines (38 loc) · 2.24 KB
Raw

TODO

  1. Ingestion
    1. Add differentiation between private and public leaderboard.
    2. Speed up: use proxies / more advanced scrapping solutions to overcome Kaggle throttling and parallelize scraping
    3. Better structure of discussions - preserve parent/children relationship between comments (currently ignored)
    4. Add kernels (code) + comments
    5. Use Meta Kaggle & Meta Kaggle Code data when possible (covers only finished, not community-based competitions)
    6. Filter out not-informative comments before ingestion (e.g. "Thanks", "Thank you", "I appreciate it", etc.)
    7. Include embedded images in all texts (currently ignored)
    8. Add full competition dataset. Maybe separate regime / index / skill
  2. Retrieval
    1. Experiment with different embedding models, potentially:
    2. Different retrieval approach: ColBERT/RAGatouille
    3. Multiple retrievers with retrieval routing / query analysis LangChain example
  3. LLM
    1. Add ollama support for LLM model
  4. UI (optional)
    1. move most of the heavyweight stuff to backend and make UI a bit more responsive
  5. Evaluation
    1. More advanced evaluation techniques
    2. Use existing libraries like RAGAS, evidently, etc.
    3. Move evaluation from jupyter notebooks to scripts / part of library
  6. Advanced assistant functionality
    1. Conversation mode - long-form chat, multiple turns
    2. Code completion conditioned by context from competition (e.g. all kernel notebooks, or best kernels, etc.)
    3. Skills & tool use
      • connect with existing Kaggle API
      • see also fastkaggle
      • add new skills (e.g. "find similar past competitions", "do EDA of the dataset", "build a baseline model", etc.)
    4. use whole platform data
  7. Tests
    1. Library level