From curated literature to a queryable knowledge graph

A Healthcare & Life Sciences business working with Cloud Combinator on AWS. The client is anonymised at their request.

The challenge had three focus areas.

At the core is Amazon Bedrock Knowledge Bases with Amazon Neptune Analytics as the graph store. When documents from the data lake gold layer are synced from Amazon S3, Bedrock automatically chunks them, uses a large language model to extract entities and relationships, generates vector embeddings with Titan Embeddings V2, and constructs the graph nodes, edges, and vector index in Neptune Analytics. This removes the need for a custom natural-language extraction pipeline, a manual schema framework, or hand-loading of the graph. Because Neptune Analytics is an unified graph and vector engine, downstream consumers can run graph traversal and semantic similarity in the same place, queried with openCypher.

The engagement delivered an automated knowledge-graph platform: a configured Bedrock Knowledge Base backed by Neptune Analytics, a Step Functions orchestration handling scheduled ingestion, validation, and the pause and resume lifecycle, full infrastructure as code, and documentation. It gives the client a graph that grows with the literature and answers both relationship and similarity queries from a single engine. Figures below reflect the design targets set in the scope of work.

• Automated extraction, no custom pipeline — Bedrock Knowledge Bases reads each document and builds the graph directly, extracting entities and relationships with a LLM rather than a hand-maintained NLP stack.
• Evidence-linked relationships — Every extracted relationship is tied back to its source paper, so the graph doubles as an evidence trail for the AI reporting that consumes it.
• Graph and vector in one engine — Neptune Analytics serves both graph traversal and semantic similarity through openCypher, removing the need for a separate vector store.
• Pause-and-resume economics — A scheduled Step Functions lifecycle keeps Neptune Analytics paused between ingestion runs, cutting its cost to a tenth of the active rate while keeping the graph ready to serve.
• Handover — Cloud Combinator delivered the platform as infrastructure as code with architecture decision records, runbooks, and a data dictionary, including an optional entity-standardisation path if naming variants fragment the graph.

With a queryable, evidence-linked knowledge graph in place, the client is ready to build its AI reporting prototype on top, and to grow the graph continuously as new literature is published without changing the underlying architecture.