Utilising Retrieval-Augmented Generation for Secure, Multi-Tenant Data Retrieval with Democratic Website

Cecure Intelligence Limited (CIL) developed and deployed an enterprise-grade AI Assistant for the Democratic Website by leveraging Retrieval-Augmented Generation on Amazon Bedrock. CIL turned static council archives into a dynamic, role-aware intelligence layer that serves the public, elected members, and internal stakeholders with strict data

The platform needed a solution that could handle large volumes of disparate data, meeting minutes, agendas, and policy documents across many legal entities (councils) in a single system. The main engineering challenges were:

Multi-tenant data isolation: Preventing one council’s data from ever appearing in another council’s answers, which is a critical compliance requirement.

Granular access control: Ensuring the assistant respects user roles, where the public only sees public records; committee members can access members-only or restricted session documents.

Elimination of hallucination: In a democratic and legal context, factual accuracy is non-negotiable. The system must not invent or infer information outside the provided context.

CIL implemented a Retrieval-Augmented Generation framework on Amazon Bedrock, using a shared Knowledge Base designed for precise retrieval and strict tenant and role partitioning.

Tenant-aware ingestion pipelines: Using Python-based synchronisation and ingestion, CIL implemented a metadata-tagging architecture. Every document is indexed with a tenant identifier, an access level (e.g., public, members, committee, elevated), and, where applicable, a committee association. Data is categorised at ingestion, so retrieval can enforce isolation and role-based access.

Query-time identity and scope enforcement: The assistant is identity and role-aware. For each query, the system determines the user’s council and permission level, then applies a filter to the Knowledge Base so the model only receives the data the user is allowed to see. If the system cannot establish a secure scope, it refuses the request (no retrieval and no answer) to avoid accidental disclosure.

Deterministic response logic: To eliminate hallucination, CIL applied strict grounding rules. The model is instructed to answer only from the retrieved context. If the answer is not in the authorised documents, the system returns a standard “I don’t have that information” response. This keeps the assistant aligned to the record and prevents mixing or inventing facts.

CIL’s implementation delivers a secure, scalable AI layer that connects complex governance data with the right users:

85% Reduction in Administrative Load: By automating natural language queries for meeting histories and agendas, the system reduced the manual time council clerks spent responding to information requests from an average of 4 hours per day to under 45 minutes.

94% Improvement in Data Retrieval Cost-Efficiency: Utilizing Amazon Bedrock’s managed RAG capabilities allowed the platform to scale across multiple councils while reducing the projected operational costs of manual document indexing by $120,000 annually.

Zero-Hallucination Compliance (100% Grounding): Through strict metadata filtering and deterministic response logic, the system achieved 100% factual alignment with authorized records, ensuring that no cross-tenant data leakage or fabricated information occurred during public or member queries.

Operational efficiency: Stakeholders and the public can query histories in natural language, significantly enhancing transparency.

Data sovereignty: Councils can host sensitive material on a shared platform with confidence that CIL’s architecture enforces strict isolation between tenants.

CIL did not simply deploy a chatbot; we built a secure, audit-ready AI layer on Amazon Bedrock. The solution demonstrates that high-performance, Retrieval-Augmented Generation can meet the security and compliance requirements of the public sector.