AI Q&A Agent For FRE Transport & Bundle Adoption: ...

Overview

How we built a conversational Q&A Agent for the FRE (First Run Experience) team at SAP, purpose-trained on Transport Management and Bundle Adoption knowledge, using Claude via SAP Hyperspace AI proxy and the Anthropic SDK.

The goal is to give application teams a fast, interactive way to understand FRE Transport and Bundle concepts before they start adoption work, without having to go through wikis, design docs, recordings, and Git repos.

Why We Built This

Onboarding a new application team to FRE adoption involves understanding a lot of context spread across multiple sources including transport lifecycle, bundle types, adoption repo setup, and Temporal workflows. A Q&A agent lets teams ask natural questions and get structured answers instantly, reducing back and forth with the FRE team.

Architecture

Key Design Decisions

1. Knowledge base as structured context: The full domain knowledge is embedded directly in the system prompt as a structured markdown document. This works well for a bounded domain with no RAG or vector store needed.

2. Smart caching with fuzzy matching: Exact match via MD5 hash gives instant replay for repeated questions. Fuzzy match via Jaccard similarity catches near duplicate phrasing. The cache persists across restarts via .qa-cache.json.

3. Continuous learning: Each interaction is stored in .qa-learned.json with up to 100 entries and a 14 day TTL, then injected back into the system prompt so the agent improves within and across sessions.

4. Hyperspace AI proxy: Uses SAP's internal Hyperspace AI desktop app as a local proxy on http://localhost:6655/anthropic so no direct API key management is needed for SAP developers.

5. Adaptive thinking: Uses thinking with type adaptive so Claude reasons through complex questions before answering, which is especially useful for multi step workflow questions.

What the Agent Knows

Transport Management

All 16 DB level transport request statuses and their UI mapping
Full CTMS flow from upload to export to import to GACD deploy
Error scenarios and validation rules on transport request creation
Config level statuses

Bundle, Transport and Adoption

Day0 (Configuration Bundle) vs Day 1 (Transport Bundle)
3 tier vs 2 tier landscape transport rules
All 3 adoptionType options (freProvidedAPI, custom, odata) with examples
CDS annotations and bundle-config-metadata.json field reference

Adoption Repo Setup

processes from fre-installation script to production ready repository
Temporal (cobalttemporal) configuration in mta.yaml
fre-client-library setup and local dev port layout

SyncMetadata Job

10 minute BTP Job Scheduler trigger
Full sync flow through BundleControllerService with MD5 change detection

Try It Yourself

Prerequisites: Node.js 18+ and SAP Hyperspace AI desktop app running

git clone https://github.tools.sap/First-Run-Experience/fre-srv/blob/rupa_qa_agent_transport/

cd fre-srv/qa-agent

npm install

export ANTHROPIC_API_KEY=

npm start

Example questions to try

What are all the transport request statuses?

How does the SyncMetadata job work?

What is the difference between freProvidedAPI and custom adoption type?

How do I configure Temporal in mta.yaml

What errors can occur when creating a transport request?

How does fre-srv trigger adoption repo workflows?

Current Scope: Trained on transport design docs, adoption template wikis, fre-srv, suite-adoption, adoption-template and client-library repos, and UI demo recordings. Answers stay grounded in the knowledge base and out of scope questions are acknowledged clearly. Local CLI only for now.

Source link