Customer support RAG

Customer support is the most common production RAG use case. The problem is well-defined (answer user questions from a knowledge base), the ROI is measurable (deflected tickets, faster resolution), and the content is usually available (help docs, past tickets, product manuals). It's also the application where the failure modes are most visible to users.

The architecture

user query
   |
route: simple vs complex
   |
[simple] →  hybrid retrieval → rerank → generation with citations
   |
[complex] → agentic RAG → multi-step retrieval → generation
   |
user receives answer + "did this help?"
   |
escalate to human if dissatisfied

The sources

Typical corpus for customer support RAG:

Weight and filter by source quality. Canonical help articles > past tickets > community posts.

The content decisions

Include past tickets?

Pros: real user language, real solutions, broad coverage.

Cons: PII, outdated information, inconsistent quality, embarrassing past answers.

Typical answer: yes but with filtering. Include only resolved tickets, strip PII, boost canonical sources above tickets.

Include forum content?

Pros: covers questions not in docs.

Cons: wrong answers, out-of-date advice, possibly contradicting official docs.

Typical answer: include with explicit labeling ("community answer, not official") and lower trust weight.

Handle multiple products/versions?

Users ask about specific products and versions. Metadata filtering:

The generation prompt

SYSTEM: You are a customer support assistant for [Company]. Answer
the user's question using only the information in the retrieved
context below. Follow these rules:
- Be concise and direct
- Cite the source of each claim with a number like [1], [2]
- If the retrieved context doesn't contain the answer, say so clearly
  and suggest escalating to human support
- Don't make assumptions about the user's account or specific setup
- Match the user's language and tone

RETRIEVED CONTEXT:
[1] [chunk 1 with source title]
[2] [chunk 2]
[3] [chunk 3]

USER: [question]

Citations are non-negotiable

Customer support answers must link to the source. Users verify. Support agents verify. Compliance requires it.

Every answer includes:

The fallback path

When the system can't answer confidently:

Never hallucinate. A confident wrong answer is much worse than "I don't know."

Personalization

When user context is available, incorporate it:

Be careful: personalized answers require accurate user data. Wrong personalization is worse than no personalization.

Handoff to human

The "escalate" path matters as much as the answer path:

Metrics that matter

Continuous improvement loop

  1. Review queries that got bad feedback
  2. Identify patterns: missing content, retrieval failures, generation problems
  3. Add missing content to the knowledge base
  4. Tune retrieval for common failure patterns
  5. Add failing queries to eval set to prevent regression

This loop is how customer support RAG gets better over time. Without it, quality stagnates.

The content governance challenge

The AI is only as good as the knowledge base. If docs are stale, incomplete, or contradictory, users get bad answers. Customer support RAG creates organizational pressure to improve documentation, which is usually a feature, not a bug.

Common mistakes

Next: Internal knowledge RAG.