Add an `assessment` object to the legal RAG search response: primary norm,
connected norms (from the citation graph references_out/in of the primary),
cross_regime, human_review_flag, a norm-level winner_margin and a short
reasoning string. The margin is computed over DISTINCT norms, so a long
article split into several chunks no longer fabricates uncertainty. The
per-result schema stays frozen — graph fields are internal (json:"-").
Also wire optional citation-graph expansion (RAG_GRAPH_EXPANSION=true,
default off): top hits pull their referenced norms into the candidate pool
via the precise edge (e.g. Art. 13 CRA -> Anhang I). Measured to add no
rank gain over the existing binding-law augmentation, with +1 Qdrant call
per search and reverse-edge fan-out risk, so it ships off-by-default as a
recall safety net. The graph EXPLAINS retrieval (assessment), it does not
expand it by default.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Benjamin Admin