In-depth analysis · About 4 min read
Octave’s Bedrock case study puts tenant isolation and retrieval ahead of the AI feature
The customer describes a shared foundation for complaint triage, with corpus-scale and productivity figures that require careful denominators.
- Source published
- Published here
- Source checked
An October 20 AWS–Octave–Cohere case study describes Reliance’s AI foundation: tenant-scoped document preparation, isolated agent execution and content screening. Complaint triage is its first vertical use case. The assistant assembles context and historical corrective actions, while a quality handler keeps the final decision. This is a customer implementation report, not a general evaluation of regulated-industry AI.
Listen to the news
Read this article aloud or keep listening to the news.
Preparing the reader…
Device voice
Speech is handled by your browser or device. AISubHub does not send the text to an audio API. Voice availability and playback with a locked screen vary by device.
Resuming may repeat a short phrase.
At a glance
- The platform’s document estate exceeds 2 billion records; the reported pipeline rate is 1.4 million documents per 24 hours per tenant workload, not 2 billion documents embedded in one day.
- Octave reports 512-dimensional embeddings meeting its target on representative queries and about 66% less vector storage than 1,536 dimensions; that is not a 66% reduction in the entire AI bill.
Representing documents involves a local quality trade-off
Reliance supports more than 650 enterprise customers and over 40 applications; the article also gives a broader Octave customer count above 14,000, a different scope. It chooses Cohere Embed v4 on Bedrock to retrieve text and images in one vector space across languages. The team tests its own quality-management corpus rather than assuming a public leaderboard resolves its needs. On representative queries, 512 dimensions meet its target while reducing vector storage by roughly 66% relative to 1,536. That result concerns this corpus and vector storage, not universal accuracy retention or total application cost. The article names in-Region availability in Virginia, Ireland and Tokyo for its requirements, while warning that model and feature coverage varies over time.
Isolation and screening act at different boundaries
The reported architecture scopes vector retrieval by tenant and runs each user session in a dedicated AgentCore microVM. Identity propagates through the execution path, with user-specific token handling and Gateway policies at tool, operation and parameter levels. This reduces reliance on each feature implementing every boundary from scratch. Content screening is separate: extracted attachments are checked before embedding, retrieved documents and tool outputs before entering context, and the generated brief before delivery. Input tagging treats attachments as data, and the document-extraction sandbox is described as having no network egress or persistence beyond the session. These are layered design controls, not evidence of zero cross-tenant disclosure or universal prompt-injection prevention. Formal checking of defined policies does not prove that the policies capture every intended business rule.
A large corpus is not one day’s completed workload
Document events pass through SNS and SQS into Step Functions, which coordinates parallel Lambda parsing, chunking and embedding. Chunks carry tenant, document type, source application, language and content-hash metadata. Batching roughly 4K-token chunks into the embedding model’s 128K context reduces request count. The case reports 28 million tokens per minute in internal testing and a sustained 1.4 million documents per 24 hours per tenant workload after batch tuning. These have different units and scopes. Neither is a general Bedrock quota or a statement that all 2 billion platform documents finished within one day. Document format, OCR, chunk count, concurrency and account capacity can change the effective onboarding rate.
The workflow supports the handler’s judgment
The complaint assistant extracts attachments, summarizes the record with Claude, searches the tenant’s OpenSearch index for similar complaints and corrective actions, scores retrieved candidates in a separate model call, and synthesizes a brief and suggested priority. A similarity match is not accepted solely because words overlap. DynamoDB caches completed briefs to avoid some repeated model work, without making all repeat requests cost-free. OpenTelemetry traces and model logs support investigation; offline evaluations compare faithfulness, retrieval precision at K and completeness with expert annotations after changes. Octave attributes 4–6 hours of preparatory work per investigation to the workflow and reports deployment time per new use case reduced by 50%. These are local customer observations, not clinical outcomes or independently controlled productivity evidence. Audit-readiness monitoring, incident narratives and predictive quality risk are described as next steps; the human quality handler still makes the consequential call.
Why it matters
AISubHub analysis: reusable retrieval, identity and execution controls can help later features inherit a reviewed foundation. They still need tenant-boundary tests, retrieval-quality checks and a human decision policy; infrastructure alone does not establish regulatory compliance.
Limits and availability
Throughput figures come from the described internal tests and pipeline tuning; productivity observations come from the reported three-month production period. All are customer-reported. We did not inspect the system or reproduce throughput, savings, isolation or injection resistance. Follow-on use cases remain plans.
Sources and evidence
Source published: Oct 20, 2026 · Source checked: Oct 21, 2026
AI-assisted editorial articles checked against linked primary sources. Vendor claims are attributed; inclusion is not endorsement.