Octave’s Reliance platform uses Amazon Bedrock services to search regulated records across tenants, with its Complaint Triage Assistant running in production.
Octave has built a shared AI foundation for Reliance, its quality-management platform for regulated industries, and is using it to power a production Complaint Triage Assistant. In an AWS post published October 20, 2026, the companies describe an architecture designed to let AI search customer records while keeping each customer’s data isolated.
Reliance supports more than 650 enterprise customers and over 40 applications, including complaint management, corrective and preventive action (CAPA) tracking, audits and supplier quality. Octave says the platform contains more than 2 billion documents, spanning formats such as PDFs, scanned forms, spreadsheets, images and email. The company says it built reusable infrastructure before developing individual AI features; that foundation has reduced deployment time per use case by 50 percent.
The design separates document preparation, retrieval and execution. Cohere Embed v4 on Amazon Bedrock creates embeddings, which are stored in a separate Amazon OpenSearch Service vector store for each tenant. Octave reports that 512-dimensional embeddings maintained its target accuracy while reducing storage by about 66 percent compared with the model’s 1,536-dimension default. Embed v4’s support for text and images in one vector space, along with multilingual capability, also helps the system retrieve across mixed media and languages.
For execution, Amazon Bedrock AgentCore Gateway applies authentication and policy before requests reach isolated microVM sessions. Amazon Bedrock Guardrails screen inputs and outputs. In the assistant’s workflow, retrieved records are checked for confidence and relevance before Anthropic Claude Sonnet generates a brief with key facts, urgency indicators, historical examples and a suggested triage priority. Octave says the system provides context; a quality handler still makes the decision.
Octave reports the assistant had been in production for three months and that its workflow absorbed work it estimates at 4–6 hours per investigation. The company says it is extending the foundation to audit-readiness monitoring, incident narratives and predictive quality-risk scoring.