TrustRAG

Standard RAG pulls every semantically similar chunk it finds, leaking confidential documents across roles and inventing unsupported claims. That's a non-starter in compliance-heavy domains.

AI systems engineer

• Pushed role and tenant filters down into the vector query, so unauthorised chunks never reach the context window.
• Mapped every synthesised sentence back to chunk IDs and source versions for a full audit trail.
• Built an evaluation harness covering 200+ permission-boundary cases, and refused to answer below a 0.78 similarity threshold.

[User Query + User JWT / Role Context]
                   │
                   ▼
       [Permission Gatekeeper] ── Validate clearance level (Public, Staff, Exec)
                   │
                   ▼
     [Hybrid Vector & Metadata Search]
   (Dense embeddings + metadata filter: tenant_id == X AND clearance <= user_clearance)
                   │
                   ▼
        [Verified Chunks + Provenance Citations]
                   │
                   ▼
    [Grounded Synthesis Engine (Hallucination Guardrails)]
                   │
                   ▼
   [Strictly Cited Answer + Source Audit Trail]

• ▸Pre-retrieval filtering: metadata filters pushed into the vector query so unauthorised chunks never enter the context window.
• ▸Provenance tracking: every synthesised sentence maps to chunk IDs, byte offsets and source document versions.
• ▸Terminal fallback: below a 0.78 cosine similarity threshold the model refuses rather than guessing.

Hardening: Separate partition keys and row-level security across vector collections to prevent cross-tenant leakage.

→100% boundary isolation across the cross-clearance leakage benchmarks, with cited answers or an explicit refusal.

• ✓100% boundary isolation in cross-clearance leakage benchmarks.
• ✓No hallucinated answers on out-of-domain queries thanks to the confidence gate.