AEGIS RAG Security Lab
A local lab for comparing keyword, semantic, and graph-based retrieval, then examining their security risks.
Scope of the work
Local document ingestion, retrieval pipelines, comparison tooling, tests, and a threat model.
Where it stands
Retrieval and threat-modeling work is documented. Adversarial tests and the initial defensive controls remain next steps.
Recorded work
completedCompleted Phase 02 threat modeling
Completed the initial AI security threat model for the RAG lab after implementing keyword, semantic, and GraphRAG retrieval paths. The phase defined assets, actors, trust boundaries, retrieval-specific attack surfaces, attack paths, and control gaps before adversarial test cases or defensive controls are added.
completedImplemented and merged Phase 01.75 local GraphRAG retrieval
Built a local GraphRAG layer with graph nodes, graph edges, concept aliases, deterministic graph construction, traversal-based retrieval, a GraphRAG pipeline, CLI runners, all-mode retrieval comparison, pytest coverage, and phase documentation. Merged the completed phase into main while preserving the phase branch as an architectural checkpoint.
completedValidated GraphRAG behavior locally
Ran the full pytest suite successfully and exercised the GraphRAG CLI against a retrieval/confidential-source-material query. The output confirmed three loaded documents, three indexed chunks, twelve graph nodes, eleven graph edges, and explainable graph retrieval details.
completedDocumented planned Phase 01.75 GraphRAG architecture
Outlined the next retrieval-architecture expansion: a local, deterministic GraphRAG path for comparing keyword retrieval, semantic retrieval, and relationship-aware retrieval before formal threat modeling begins.
completedMerged Phase 01 and Phase 01.5 retrieval pipeline
Integrated the baseline keyword RAG pipeline and Phase 01.5 semantic retrieval path into main, including local document ingestion, chunking, keyword retrieval, deterministic local embeddings, in-memory vector retrieval, comparison CLI tooling, pytest coverage, and phase documentation.
Next steps
- Review the Phase 02 documentation state after the connector truncation/debugging work
- Convert threat-model findings into adversarial retrieval test cases
- Define retrieval-layer controls for source trust, context handling, and graph relationship abuse
- Keep implemented controls separate from planned controls in public documentation