RAG for Support: Fresh Answers Without Model Retraining

yaml
pipeline:
  name: support-rag-pipeline
  owners:
    support_owner: "manager-emea-support@company.com"
    docops_owner: "doc-operations@company.com"
    legal_contact: "privacy@company.com"
  regions:
    primary: eu-west-1
    backup: us-east-1
  sources:
    - id: zendesk_guide
      type: zendesk
      url: https://support.company.com/hc
      rbac_group: "agents"
      freshness:
        recrawl_cron: "*/15 * * * *"   # every 15 minutes for hotfixes
        slo_days: 2
      metadata:
        tags: ["macro", "howto"]
        locales: ["en-US", "de-DE", "fr-FR"]
    - id: confluence_runbooks
      type: confluence
      space: SUP
      rbac_group: "internal_only"
      freshness:
        recrawl_cron: "0 * * * *"      # hourly
        slo_days: 7
      metadata:
        tags: ["runbook", "escalation"]
        locales: ["en-US"]
    - id: release_notes
      type: github
      repo: company/product-release-notes
      path: /releases
      rbac_group: "agents"
      freshness:
        recrawl_cron: "*/10 * * * *"   # every 10 minutes during launch days
        slo_days: 1
      metadata:
        tags: ["changelog", "versioned"]
  processing:
    pii_redaction:
      enabled: true
      policies: ["email", "phone", "iban"]
    chunking:
      strategy: semantic
      target_tokens: 400
      overlap_tokens: 60
    embeddings:
      model: "text-embedding-3-large"
      normalize: true
    reranker:
      model: "cross-encoder-ms-marco"
      top_k: 8
    recency_boost:
      half_life_days: 14
      max_boost: 1.6
    synonyms:
      lexicon: ["2FA:two-factor","tenant:workspace","billing toggle:billing feature flag"]
  retrieval:
    filters:
      - key: locale
        value_from: ticket.locale
      - key: product
        value_from: ticket.product
      - key: version
        strategy: nearest_effective_date
    citations:
      required: 2
      style: inline_links
    thresholds:
      confidence_auto_suggest: 0.78
      coverage_min_chunks: 2
      escalate_below_confidence: true
  hitl:
    approval_flow:
      surfaces: ["zendesk_sidebar"]
      approvers:
        - role: "senior_agent"
          limit: "< 0.78 confidence"
        - role: "sme"
          limit: "new-topic or policy-tagged"
      feedback:
        channels: ["slack:#kb-approvals", "teams:Support Ops"]
        fields: ["good_answer","needs_edit","wrong_source","low_confidence"]
  governance:
    rbac:
      groups:
        agents: ["zendesk"]
        internal_only: ["zendesk_internal","servicenow"]
    logging:
      prompt_logging: true
      response_logging: true
      redact_values: ["email","iban"]
      retention_days: 365
    residency:
      region_lock: true
    training:
      never_train_on_client_data: true
  telemetry:
    kpis:
      deflection_rate_target: 0.22
      aht_target_seconds: 310
      csat_delta_target_points: 3
    dashboards:
      weekly_brief: "slack:#support-leadership"
      export: "s3://support-ai-telemetry/rag/"

{
  "title": "RAG for Support: Fresh Answers Without Model Retraining",
  "published_date": "2025-12-10",
  "author": {
    "name": "Alex Rivera",
    "role": "Director of AI Experiences",
    "entity": "DeepSpeed AI"
  },
  "core_concept": "AI Copilots and Workflow Assistants",
  "key_takeaways": [
    "RAG keeps answers fresh by retrieving the latest approved content at runtime—no heavy retraining cycles.",
    "A governed retrieval pipeline gives you citations, recency boosts, and RBAC so agents trust suggestions.",
    "30-day motion: Week 1 audit and voice tuning; Weeks 2–3 retrieval pipeline and prototype; Week 4 analytics and rollout plan.",
    "Human-in-the-loop gates and confidence thresholds protect CSAT while reducing handle time.",
    "Deploy in your stack: Zendesk or ServiceNow, Slack/Teams, and a vector store—no data leaves your control, with full audit trails."
  ],
  "faq": [
    {
      "question": "How does RAG handle conflicting content across locales?",
      "answer": "We filter by locale and product before generation and boost content with the most recent effective date. Conflicts are flagged to content owners via Slack for review."
    },
    {
      "question": "Can we keep internal runbooks from appearing in external answers?",
      "answer": "Yes. RBAC tags on the source collections ensure internal-only documents are retrieved only for agent surfaces, never for end-user deflection."
    },
    {
      "question": "What if confidence is low?",
      "answer": "Below your threshold, the copilot requires agent approval or escalates to an SME. Low-confidence topics are logged so DocOps can patch gaps quickly."
    },
    {
      "question": "Do we need to fine-tune the LLM later?",
      "answer": "Most teams don’t. We tune prompts for brand voice and rely on retrieval freshness. If you later pursue fine-tuning, it’s narrow and optional."
    }
  ],
  "business_impact_evidence": {
    "organization_profile": "B2B SaaS, 450-person support org, Zendesk + Confluence, multi-locale help center",
    "before_state": "Weekly releases left macros and help-center content stale; agents relied on memory. AHT spiked during launches and deflection lagged.",
    "after_state": "RAG copilot surfaced fresh, cited answers with agent approvals. Hotfixes reached the index in 15–20 minutes; low-confidence answers gated to SMEs.",
    "metrics": [
      "AHT reduced from 378s to 310s on billing/auth queues (-18%)",
      "Self-serve deflection increased from 16% to 40% (+24 pts)",
      "Reopen rate down from 11% to 6% (-5 pts)",
      "Approx. 2,100 agent-hours returned per quarter"
    ],
    "governance": "Legal and Security approved due to prompt/response logging with redaction, RBAC over source collections, EU data residency enforcement, human-in-the-loop approvals, and a policy that models are never trained on client data."
  },
  "summary": "Support leaders: use RAG to keep answers current without retraining models. 30‑day plan: knowledge audit, retrieval pipeline, prototype, and analytics."
}