ContextGuard

State-contracted verification and explainable retrieval gating for multi-turn RAG.

What it is - A constraint-first verification engine: plan → retrieve → gate → judge → aggregate → report/context pack. - Contracts and types: StateSpec, Claim, Chunk/Provenance, ReasonCode. - Outputs: verdict report (primary), context pack (secondary), trace DAG (explainability).

Why it exists - Similarity ≠ relevance under constraints. Wrong-year/entity/source chunks cause confident but wrong answers. - ContextGuard hard-gates evidence using state constraints and reason codes, making failures visible and auditable.

Key capabilities - Planner (support + counter), hard gating (entity/time/source policy/diversity), judges (rule/LLM/NLI), aggregation with source priority, numeric/time safeguards, trace DAG, rich reports. - Adapters: LLM providers (OpenAI + budget/retry), retrievers (LangChain, LlamaIndex, Chroma, Qdrant), stores (SQLite, S3), async pipeline, generation scaffold.

Who it’s for - Teams building RAG/agent systems that must be constraint-aware, auditable, and reproducible.