RAG knowledge bases
Amazon Bedrock Knowledge Bases over ingested corpora: contracts, SOPs, product data, Smart OCR-extracted documents, with citation and access control.
- Bedrock KBs
- Citations
- Access control
Data & AI · GenAI, RAG & Agents
Copilots that know your business.
RAG knowledge bases and enterprise copilots on Amazon Bedrock; SAP Joule enablement where the data lives in SAP.
Where the answer needs to be grounded in real enterprise data (SAP masters, contracts, historical decisions) Phoenix builds RAG and agentic solutions on Amazon Bedrock. The retrieval side reuses the document AI and data-platform work; the generation side sits on foundation models chosen per use case.
Where the workload lives in SAP, SAP Joule is enabled directly. Where it doesn't, Bedrock stays the platform of record. In both cases we're wiring copilots to the same data the business already trusts.
What's in scope.
Enterprise copilots
Domain copilots that pull from RAG and call tools (SAP transactions, ticketing, CRM lookup) under a guardrails layer for permissions and audit.
- Tool-calling
- Guardrails
- Audit trail
Agentic workflows
Multi-step agents built on Bedrock Agents or on custom orchestration where control matters: chain reasoning, evaluations, and human-in-the-loop checkpoints.
- Bedrock Agents
- HITL
- Evals
SAP Joule enablement
Where the workload sits in S/4HANA or SuccessFactors, SAP Joule is turned on and integrated: a single conversational surface across SAP and non-SAP data.
- SAP Joule
- S/4HANA
- SuccessFactors
The approach.
-
01
Frame
Real use case, target user, acceptance criteria: not a PoC-for-PoC-sake.
-
02
Ground
Corpus curated, RAG ingested, retrieval tuned before any model choice is finalised.
-
03
Ship
Copilot or agent goes to a defined user group with guardrails and telemetry from day one.
-
04
Evolve
Evaluation loop, model swaps, tool expansion: production-grade lifecycle.
Named deliverables.
Every engagement lands specific artefacts, not slides.
- RAG knowledge base on Amazon Bedrock, grounded in your corpus
- Copilot UI (web or in-app) with citations and permission enforcement
- Agent tool-integrations to SAP or other systems, where in scope
- Guardrails and audit trail for every model call
- Evaluation harness: accuracy, hallucination rate, latency, cost per query
- SAP Joule configuration where applicable, integrated with existing SAP roles
Frequently asked
Which foundation models does Phoenix work with?
Model choice is per use case. Amazon Bedrock is the default surface (Anthropic Claude, Meta Llama, Amazon Nova and others available through Bedrock) with the option to fall back to hosted OpenAI where a customer requires it. SAP Joule is the vehicle when the workload is inside S/4HANA or SuccessFactors.
How do you stop the copilot from hallucinating?
Retrieval-Augmented Generation grounded in a curated corpus, mandatory citations on every answer, evaluation harness measuring hallucination rate against a labelled set, and guardrails at the prompt and response layer. Answers without citations are treated as failures.
Can copilots actually take actions in SAP or just answer questions?
Both. Read-only lookups over SAP are the safer starting point. Agentic workflows can call SAP transactions or trigger BTP APIs under guardrails, an approval matrix and full audit: the same governance that a human user would operate under.
How do you handle access control on retrieved documents?
Retrieval respects the same row-level and column-level access model the source system uses: a user asking a copilot only ever sees documents they're already permitted to see, and citations respect that permission. Access enforcement happens at retrieval time, not just at display time.
How do you measure whether a copilot is worth keeping in production?
Every deployment lands with an evaluation harness: accuracy on a labelled set, hallucination rate, latency, cost per query, and, most importantly, a business KPI (deflection rate, cycle-time reduction, first-response time). If the KPI doesn't move, the copilot doesn't stay.
The earliest conversations are usually the most useful.
Whether you're scoping an SAP move to cloud, restarting a stalled programme, or just trying to figure out where data and AI fit, start with a conversation.
