New

Free VAPT consultation for new enterprise clients — Book your security assessment

Case study · Enterprise

Private RAG assistant for customer support

A services enterprise with thousands of support tickets monthly wanted faster first responses without sending customer data to public model providers. We built a grounded assistant over their knowledge base with role-aware retrieval, cited answers, and human approval before any customer-facing reply.

Representative engagement · details anonymized under NDA

Project overview

Client
Enterprise services company (name withheld under NDA)
Industry
Enterprise
Services
RAG Development, AI Agents, AI Automation

The problem

Agents searched across wikis, tickets, and PDFs manually; answers varied by agent. Compliance prohibited customer data in public LLM APIs, ruling out off-the-shelf chatbots.

Requirements

  • Answers grounded strictly in approved internal content with citations
  • Role-based access — agents see only permitted documents
  • Human approval before customer-facing responses
  • Deployment inside the company's cloud boundary
  • Measurable quality gates before launch

Key challenges

  • Messy corpus: duplicate wikis, stale runbooks, conflicting procedures
  • Strict data residency and PII handling requirements
  • Support leads distrusted AI drafts after a bad pilot elsewhere

Our solution

  • Curated and deduplicated the corpus with owners, then built metadata-aware ingestion
  • Implemented hybrid retrieval with reranking plus pre-retrieval permission filters
  • Built a golden eval set from real tickets and gated releases on faithfulness scores
  • Deployed open models in the client's VPC with PII redaction and full audit logs

Technology stack

pgvectorHybrid retrieval + rerankerOpen models (VPC)Function callingEval harness

Security implementation

  • Pre-retrieval tenant and role filtering — restricted content never reaches the model
  • PII redaction on inputs and outputs
  • Immutable audit log of prompts, retrieved context, and actions
  • Human-in-the-loop approval for all customer-facing replies

Performance considerations

Cached embeddings and rerank-only-top-k kept p95 answer latency within support-workflow budgets; nightly evals tracked drift.

Results

  • Assistant drafts adopted by support leads after eval-gated pilot
  • Every answer traceable to source passages for audit
  • Eval harness retained by the client for ongoing model updates

Lessons learned

  • Corpus curation beats model upgrades for answer quality
  • Permission filtering must happen before retrieval, not after generation
  • Trust comes from evals and citations, not demo fluency

Have a similar challenge?

Tell our Gurgaon team about your project — reply within one business day.

Start the conversation