This directory is the implementation-grounded methodology reference for Luxia's hybrid corrective RAG fact verification pipeline.
This methodology documents how the current system works in code, not how a generic RAG system should work.
Primary implementation sources:
socket-hub/app/main.pysocket-hub/app/sockets/manager.pycontrol-plane-api/app/routers/v1.pydispatcher/app/main.pyworker/app/main.pyworker/app/services/corrective/pipeline/__init__.pyworker/app/services/retrieval/*worker/app/services/vdb/*worker/app/services/kg/*worker/app/services/ranking/*worker/app/services/verdict/*frontend/app/client/page.tsx
- 01-system-architecture.md
- 02-pipeline-stage-decomposition.md
- 03-retrieval-and-ingestion.md
- 04-ranking-trust-and-corrective-loop.md
- 05-verdict-synthesis.md
- 06-failure-modes-and-tradeoffs.md
- 07-system-dataflow-and-innovations.md
- figures-highlevel/README.md (recommended final paper set)
- figures/README.md (deep technical set)
Each component section uses the same six-part schema:
- Functional role in the system
- Technical mechanism (internal operation)
- Inputs and outputs
- Interaction with other components
- Why it is necessary in this hybrid pipeline
- Failure points and design trade-offs
VDBmeans Pinecone vector storage and retrieval path.KGmeans Neo4j relation graph path.Adaptive trustmeans coverage/diversity/agreement-based sufficiency gating.Corrective loopmeans incremental search -> scrape -> extract -> ingest -> re-rank.
Each major methodology file includes Mermaid diagrams plus prose equivalents. Diagrams are explanatory overlays; prose is complete and authoritative on its own.
The figures/ subdirectory provides a publication-ready figure pack (F01-F56) with dual Mermaid plus camera-ready figure specifications.
The figures-highlevel/ subdirectory provides the final publication-facing 20-figure set (H01-H20).
- Health-only runtime scope: claims are processed in health mode; there is no active multi-domain routing.
- No runtime APIs are changed by this documentation set.
- Existing docs under
docs/are treated as non-authoritative unless revalidated against code. - Terminology baseline:
hybrid corrective RAG fact verification pipeline.
Last verified against code: March 10, 2026