Capstone: Enterprise RAG Platform

Build a multi-tenant retrieval platform that ingests private enterprise documents and produces answers with exact, authorized citations.

System scope

• connectors for object storage and one document system;
• parsing for text, tables, and page layout;
• versioned chunking, embeddings, and hybrid retrieval;
• metadata and access-control filtering before retrieval;
• reranking, context assembly, citation mapping, and abstention;
• incremental updates, deletion, and index reconciliation; and
• offline evaluation plus production tracing.

Loading diagram...

Required evidence

Compare dense and hybrid retrieval, two chunking strategies, and reranking on a fixed dataset. Report retrieval recall, citation precision, grounded answer rate, latency, and cost.

Demonstrate that one tenant cannot retrieve another tenant's content. Update a source document, remove another, and prove that the serving index converges.

Failure tests

• a retrieved document contains prompt injection;
• the embedding service is unavailable;
• an index is stale during a release;
• no authorized source supports the answer; and
• a deletion request arrives while indexing is running.

Completion criteria

The platform passes when every answer maps to permitted source spans or abstains, index lifecycle is observable, and the complete release can roll back.