1 comments

  • sumoaps 2 hours ago

    Author here. This is a minimal (~1k LOC) Python implementation showing how to make retrieval strategy a learnable decision in RAG/agent systems.

    Key ideas: - Router extracts query features (digit_ratio, rare_ratio, etc.) - Chooses between keyword, vector, or hybrid retrieval - Feedback loop updates weights based on evaluation scores - All runs logged to SQLite (or Databricks Lakebase)

    The repo runs on CPU with no external API calls. Happy to answer questions about the design or production considerations.