ai-enabled-hr-talent-automation

Retrieval and Grounding Policy

Title: Retrieval and Grounding Policy Version: 1.0 Owner: [TENANT_CONFIGURATION_REQUIRED — AI Governance Lead] Status: Draft Last reviewed: 2026-09-07 Next review: [TENANT_CONFIGURATION_REQUIRED] Reviewers: AI Governance, Security

Purpose and scope

Defines how retrieval is executed and how grounding/citation is enforced, building on rag-architecture.md.

Retrieval pipeline

  1. Metadata ACL filtering — applied first, before any similarity computation: tenant, classification, access policy, effective/expiry date.
  2. Hybrid search — combines vector similarity with keyword/BM25 search to catch exact-term policy references (e.g., clause numbers) that pure embeddings may miss.
  3. Reranking — a secondary, more precise relevance pass over the top-N hybrid results.
  4. Minimum sufficient context selection — select the smallest set of chunks that answers the question, avoiding over-stuffing the prompt with irrelevant context (cost and accuracy benefit).
  5. Citation attachment — every claim in the final answer must map to a specific chunk (document_id + heading_path + page_or_clause_ref).
  6. No-answer fallback — if retrieval relevance is below the configured threshold (default 0.65 — see ai-guardrails-policy.md), the system responds that it cannot find a grounded answer and suggests contacting HR directly, rather than guessing.

Grounding enforcement

The model is instructed (see hr-orchestrator-system-prompt.md) to answer only from retrieved chunks for policy questions; general parametric knowledge is not used for policy-specific claims. Output validation checks that every citation reference in the answer corresponds to an actually-retrieved chunk ID (prevents fabricated citations).

Handling conflicting sources

If multiple retrieved chunks conflict (e.g., an outdated and a current policy version both indexed during a transition), the chunk with the latest effective_date not yet superseded takes precedence; the answer notes if a recent policy change may affect the response.

Configurable items

Relevance threshold, hybrid search weighting, reranking model choice, and max context tokens are configurable via config/schemas/rag-config.schema.json.

Change control

Version Date Author Change
1.0 2026-09-07 Documentation package generation Initial creation
This site is open source. Improve this page.