Similarity blurred ownership
Boilerplate language made unrelated records look semantically relevant.
FIELD NOTES / HEALTHCARE AI & GRAPHRAG
Clinical evidence.
Connected before AI answers.
NoteGraph turns near-identical clinical notes into a governed knowledge graph, then combines vector anchoring with explicit patient and encounter relationships to produce source-linked answers clinicians can inspect.
CLINICAL AI CASE STUDY
NoteGraph
Relationship-aware retrieval.
Evidence-support boundary.
NoteGraph / Encounter-scoped chatSOURCE-GROUNDED EXPERIENCE
ANSWER + PROVENANCEThe evidence is part of the answer.
Citations resolve to authorized source notes, while trace IDs preserve the retrieval and model configuration behind the response.
Follow the evidence ↗AuthorizeAnchorTraverseRerankAnswer & cite
DOMAIN
Clinical documentation
WHAT WE BUILT
Governed GraphRAG platform
OUR ROLE
AI, cloud & product engineering
01 / THE RETRIEVAL PROBLEM
The notes look alike.
The patients do not.
Templated OBGYN notes can be lexically near-identical across unrelated patients. Plain keyword or vector search may retrieve a persuasive-looking passage that belongs to the wrong encounter.
That is not a ranking inconvenience. In a clinical workflow, it is the wrong evidence attached to the right-sounding answer.
NoteGraph treats patient, encounter, finding, and referenced-visit relationships as retrieval evidence—before the model sees context.
Answers needed an evidence trail
Clinicians required the note, encounter, evidence span, and provenance—not an opaque summary.
Safety had to be measurable
Grounding, refusal, authorization, and retrieval quality needed explicit tests and operational signals.
02 / INSIDE THE PRODUCT
Answer the question.
Open the evidence.
The clinical experience keeps the reasoning path visible: encounter-scoped conversations, source-linked citations, and an evaluation surface for retrieval and generation quality.
Connected evidence, not a plausible document match.
The clinician asks across a patient’s record. NoteGraph returns a bounded answer, source chips, retrieval trace details, and a clear statement when evidence is limited.
THE QUESTION-TO-EVIDENCE LOOP
Authenticate→Rewrite→Vector anchor→Graph traverse→Rerank→Guard→Stream & cite
03 / THE RETRIEVAL BREAKTHROUGH
Similarity finds the doorway.
Relationships find the record.
The query first locates semantically relevant anchor nodes. Bounded openCypher traversal then expands only through approved clinical relationships, and a policy reranks the result by relevance, graph distance, recency, and evidence quality.
Look under the hood ↘01
Authorize the retrieval boundary
Identity, role, tenant, request ownership, and source permissions are checked before context assembly.
Find candidate anchors
Vector similarity identifies likely notes, findings, or encounters without deciding clinical belonging.
Traverse approved edges
Patient, encounter, finding, and referenced-visit edges constrain what can join the answer context.
02
Synthesize under a grounding contract
The model can use only the retrieved, authorized evidence and must surface insufficiency or contradiction.
03
Return an inspectable final message
Citations, trace ID, and configuration versions travel with the completed answer.
One patient journey, expressed as evidence paths.
Clinical relationships stop near-duplicate text from becoming cross-patient context.
Patient→Encounter→Note→Finding↗Prior visit
Evidence support—not autonomous diagnosis.
NoteGraph answers from retrieved, authorized documentation, cites the source, and states when the record is insufficient. Diagnosis and treatment decisions remain outside the product boundary.
04 / THE TECHNOLOGY & THE REASON
AWS services with
clear jobs and clear boundaries.
The stack separates graph evidence, transactional product data, transient connection state, and immutable provenance—so each store and workload can be secured, scaled, audited, and recovered independently.
GRAPH + VECTOR RETRIEVAL
Neptune Analytics
Runs semantic anchor search and bounded openCypher traversal against one governed clinical graph.
Why one graph store?
Similarity and relationship expansion operate on the same governed entities instead of drifting across disconnected databases.
MODEL RUNTIME
Amazon Bedrock
Provides embeddings, schema-constrained extraction, tiered Claude inference, streaming, and guardrails.
Why tiered routing?
High-complexity synthesis uses the primary model while helper and fallback paths stay explicit, versioned, and measurable.
DURABLE INGESTION
Step Functions + Fargate
Coordinates validation, OCR, normalization, de-identification, graph construction, and quality gates.
Why durable orchestration?
Long-running clinical data work becomes replayable and observable without hiding failures inside request-time functions.
PROVENANCE
Amazon S3
Raw, quarantined, and curated zones retain immutable versions, manifests, graph records, and evaluation datasets.
Why keep curated records?
The graph can be rebuilt deterministically; corrupt state becomes a recovery procedure rather than a data-loss event.
PRODUCT STATE
Aurora PostgreSQL
Owns users, conversations, messages, and configuration separately from the clinical graph.
Why physical separation?
Clinical evidence and chat-product state receive different authorization, retention, backup, and blast-radius controls.
IDENTITY + DELIVERY
Cognito + API Gateway
Federated identity, MFA, JWT authorization, HTTP APIs, and ordered WebSocket answer streams form the product edge.
Why encounter-scoped access?
The right to ask a question never implies the right to open every source note; policy is enforced on both paths.
Validate & quarantine
Version, checksum, classify, deduplicate, and route failures to replayable queues.
Normalize & extract
OCR scanned notes and produce schema-constrained clinical entities and relationships.
Load governed evidence
Publish only after node, edge, orphan, and schema quality gates pass.
Anchor & traverse
Combine semantic relevance with explicit clinical connectivity.
Stream & trace
Return bounded answers with citations, trace IDs, and configuration versions.
The supplied interface presents an evaluation dashboard with representative project metrics. Formal performance and regulatory acceptance remain governed by the client-approved golden dataset, target thresholds, environment, and applicable compliance review.
Explore the documented platform stack
- Graph & retrieval
- Neptune Analytics · openCypher · vector anchors · bounded traversal · reranking policy
- AI runtime
- Amazon Bedrock · Claude tiers · embeddings · extraction · Guardrails · versioned prompts
- Ingestion
- S3 · EventBridge · SQS/DLQ · Step Functions · Textract · ECS Fargate
- Application
- React SPA · Amplify Hosting · API Gateway HTTP/WebSocket · Lambda
- Data
- Aurora PostgreSQL + RDS Proxy · DynamoDB TTL state · S3 provenance
- Security & operations
- Cognito · WAF · IAM · KMS · Secrets Manager · CloudWatch · X-Ray · CloudTrail · Config · Security Hub · Macie
Clinical and compliance boundary
Cloud architecture does not itself confer regulatory compliance. Production PHI processing requires the applicable agreements, service-and-region review, threat model, approved operational controls, retention policy, and client governance. NoteGraph is positioned as evidence-support software—not autonomous clinical decision software.
Technology names identify the documented implementation, not partnerships or endorsements.
05 / QUALITY AS A PRODUCT SURFACE
“Looks good” is not
an acceptance test.
A golden evaluation dataset turns retrieval and answer quality into release evidence. Operators can inspect anchor precision, connected-context recall, grounded answers, unsupported claims, fallback use, latency, and denied queries by run.
The important move is organizational: retrieval, generation, experience, and security are measured together, because a clinically useful answer must pass all four.
RETRIEVAL
Anchor precision · connected-context recall
GENERATION
Grounding · unsupported claims · refusal quality
EXPERIENCE
Streaming health · time to first token
SECURITY
Denied queries · unusual source access
Insufficient evidence is a valid answer
The assistant must state when authorized notes do not support the requested conclusion.
Every source open is authorized again
A citation is not a bypass; source inspection applies the same role and tenant policy.
Fallback use remains visible
Retries and model-route changes are recorded rather than silently changing behavior.
The graph remains rebuildable
Curated records and manifests preserve a deterministic recovery path.
A CONVERSATION ABOUT YOUR KNOWLEDGE WORKFLOW
Where does your RAG system
lose the right context?
Bring one regulated or relationship-heavy document workflow. In 20 minutes, we’ll map the evidence model, authorization boundary, evaluation contract, and a practical engineering next step.
20 minutes · Your use case, constraints, and next step
20 minutes · Your use case, constraints, and next step