AI

RAG Retrieval Strategy

Improve documentation RAG through measured metadata scoping, hybrid retrieval, and reranking while keeping vector retrieval as the default and allowing graph retrieval only through a later evidence-gated spike.

State
Done
Progress
100%
Steps
37 / 37
Started
2026-10-08
Finished
2026-10-08
Module
LaraAI

Tasks.

  • Task 1: Versioned RAG evaluation dataset and loader6/6
  • Task 2: Retrieval evaluation runner and baseline report7/7
  • Task 3: Stable chunk metadata and explicit vector retrieval factory7/7
  • Task 4: Feature-flagged Elasticsearch hybrid retrieval7/7
  • Task 5: Optional bounded reranking5/5
  • Task 6: Graph retrieval decision checkpoint5/5

Read the full plan on GitHub