Built a fully serverless RAG pipeline on AWS that supplements the support workbench with historical ticket context

A fully serverless ingestion and retrieval pipeline on AWS for past solved tickets. S3-triggered Lambdas chunk and embed each resolved ticket into Amazon OpenSearch Serverless (vector engine), with retrieval exposed to the rest of the suite through Lambda functions behind API Gateway.

Retrieval runs two-tier: deterministic metadata matching first — looking for a “golden ticket” that exactly matches the new one — then falling back to semantic search across past solved tickets when no precise match exists. On top of that sits 17 category-specific lookup strategies, so the retrieval logic for a billing dispute looks nothing like the retrieval logic for a shipping delay.

The retrieved context is synthesized into actionable engineer notes that surface alongside the AI's first response draft, so the engineer sees not just what to say but what was done last time.

I designed the two-tier retrieval approach, authored the 17 category-specific lookup strategies, and built the S3-triggered ingestion Lambdas that turn past tickets into a searchable knowledge base in OpenSearch Serverless.

I also built the retrieval Lambdas and their API Gateway surface, so the rest of the suite — the First Response Agent and the Zendesk Chat App — could query historical context through a single, versioned interface.