A RAG knowledge assistant, built on AWS

A SaaS B2B business working with Cloud Combinator on AWS. The client is anonymised at their request.

The challenge had three focus areas, each concerned with giving the client accurate answers from its own material without blurring the line between audiences.

Retrieval-Augmented Generation means the assistant does not answer from the language model alone. When a question arrives, the system first retrieves the most relevant passages from the client's own documents and then asks the model to compose an answer grounded in that retrieved material. This keeps responses accurate to the client's content and reduces the risk of the model inventing an answer. The generative reasoning is provided by Claude models running on Amazon Bedrock, so no model infrastructure has to be managed directly.

The engagement delivered a working, serverless RAG knowledge assistant that answers customer and admin questions from the client's own documents, with the two audiences kept fully separated and new content available for retrieval shortly after upload.

• Grounded, audience-aware answers — Customer and admin queries are each routed to a dedicated Bedrock Agent and Knowledge Base, so responses are drawn from the right body of content and the two audiences never overlap.
• A direct ingestion pipeline — Uploaded documents are tagged with metadata, stored in a single secured S3 bucket, and ingested straight into the correct Knowledge Base, making them usable for query resolution almost immediately.
• Serverless and fault-tolerant — Stateless Lambda handlers scale with demand and carry retry logic to absorb transient failures, so the assistant stays responsive under bursty, real-time load.
• Secure by design — Services communicate inside an Amazon VPC, documents are logically segregated by metadata and access controls in one secured bucket, and DynamoDB provides fast, serverless session and activity tracking.
• Handover — The client received the deployed RAG workflow spanning ingestion, retrieval, and query routing, ready to grow its document corpus behind a stable set of APIs.

With a grounded, audience-aware assistant in place, the client has a foundation that scales with its knowledge base: as more documents are added, the same three endpoints keep serving accurate, separated answers, and the serverless architecture grows with demand rather than ahead of it.