ai-enabled-hr-talent-automation

RAG Architecture

Title: RAG Architecture

Version: 1.0

Owner: [TENANT_CONFIGURATION_REQUIRED — AI Governance Lead]

Status: Draft

Last reviewed: 2026-09-07

Next review: [TENANT_CONFIGURATION_REQUIRED]

Reviewers: Architecture, Security, AI Governance

Purpose and scope

Details the RAG architecture within the boundary set by ADR-003: approved retrieval content only, never a system of record.

Architecture

flowchart LR
    subgraph Ingestion
        D1[Approved source documents: policy, process, JD reference] --> D2[Chunking - see rag-ingestion-and-chunking.md]
        D2 --> D3[Metadata tagging: tenant, classification, access policy, dates]
        D3 --> D4[Embedding generation]
        D4 --> D5[(Vector Store)]
    end
    subgraph Retrieval
        Q1[User/agent question] --> Q2[Metadata ACL filter - BEFORE similarity search]
        Q2 --> Q3[Hybrid search: vector + keyword]
        Q3 --> D5
        D5 --> Q4[Reranking]
        Q4 --> Q5[Minimum sufficient context selection]
        Q5 --> Q6{Sufficient grounding?}
        Q6 -- No --> Q7[No-answer fallback]
        Q6 -- Yes --> Q8[Answer + citations]
    end

Key design rules

1. Metadata ACL filtering happens before similarity search, not after — a document a user/tenant is not entitled to see is never scored or returned, even partially (see retrieval-and-grounding-policy.md).
2. Never index raw candidate identity documents or confidential interview feedback by default (see ADR-003, privacy-and-pii-handling.md).
3. Every answer must cite the specific chunks used; if sufficient grounding isn’t found, the system returns a no-answer fallback rather than a hallucinated answer.
4. The vector store is fully rebuildable from rag_document/source content — treat the index as a derived cache, not a backup-critical store (see resilience-and-disaster-recovery.md).

Tenant isolation in retrieval

Each tenant’s approved content is either stored in a tenant-scoped namespace/index, or in a shared index with mandatory tenant_id metadata filtering enforced server-side before any vector search executes — never left to client-supplied filters alone.

Cross-references

rag-ingestion-and-chunking.md · retrieval-and-grounding-policy.md · rag-config.schema.json · rag.default.yaml

Change control

Version	Date	Author	Change
1.0	2026-09-07	Documentation package generation	Initial creation

This site is open source. Improve this page.