agentforce-reliable-rag-agent

Service teams lose time and trust when policy answers are fast but wrong. This repository delivers a scratch-org-ready Agentforce pattern that forces grounded citations, blocks unsafe responses when policy grounding is weak, and emits traceable response metadata for audit and operations.

Measured Results

Generated from scripts/run-tests.sh and scripts/eval/run_eval.sh into docs/benchmarks.json:

• Apex coverage (CI gate): 90.0%
• Citation coverage: 94.0%
• Unsupported-claim rate: 4.5%
• Grounded-answer rate: 94.0%

How it works

• Prompt Builder template (PolicyAnswer) defines tone, refusal behavior, and citation contract.
• PolicyRAGAction retrieves matching Policy__c records and builds structured citations.
• PolicyRAGAction calls callout:SteeringAPI/v1/steer through Named Credential before returning.
• Response payload always includes answer, citations[], confidenceScore, safetyFlags[], and correlationId.

Deploy

# Authenticate your Dev Hub once per machine
sf org login sfdx-url --sfdx-url-file ./SFDX_AUTH_URL.txt --alias devhub --set-default-dev-hub

# Create, deploy, assign perm set, and seed sample policy records
bash scripts/setup-scratch-org.sh agentforce-rag

# Run Apex tests with coverage gating (fails below 85%)
bash scripts/run-tests.sh agentforce-rag

# Run offline eval and regenerate docs/benchmarks.json + docs/benchmarks.md
bash scripts/eval/run_eval.sh

# Delete scratch org
bash scripts/teardown-scratch-org.sh agentforce-rag

Security model

• Named Credentials: SteeringAPI isolates the Steering endpoint; use org-managed secret strategy when moving beyond NoAuthentication.
• Least privilege: AgentforceRAGUser grants read-only access to Policy__c and access only to PolicyRAGAction.
• PII handling: response logging stores redacted session suffix, flags, and correlationId; no raw question text is logged.
• Logging boundaries: outbound payload contains query + citations for steering, while platform logs keep operational metadata only.

Documentation

• Architecture: docs/architecture.md
• Threat model: docs/threat-model.md
• Benchmarks: docs/benchmarks.md
• Developer handoff checklist: docs/handoff-checklist.md