Skip to content

copilot-instructions.md

Latest commit

 

History

History
62 lines (46 loc) · 3.12 KB

copilot-instructions.md

File metadata and controls

62 lines (46 loc) · 3.12 KB
Raw

SAP CodeJam: Code-Based AI Agents

Hands-on workshop for building multi-agent AI systems on SAP BTP using CrewAI and LiteLLM connected to SAP Generative AI Hub (AI Core). The scenario: investigate an art heist by orchestrating specialized agents that analyse evidence, appraise stolen items, and identify the culprit.

Architecture

Three-agent sequential crew in project/Python/:

Agent Tool Purpose
Appraiser call_rpt1() Predict item categories & insurance values via SAP-RPT-1 ML model
Evidence Analyst call_grounding_service() RAG queries over evidence documents via SAP Grounding Service
Lead Detective (none) Synthesise findings and name the culprit

Key files in project/Python/solution/:

  • investigator_crew.py — @CrewBase class; agents and tasks defined with @agent, @task, @crew decorators; tools use @tool("name")
  • main.py — entry point; calls crew().kickoff(inputs={...})
  • rpt_client.py — OAuth2 client for SAP-RPT-1 (client-credentials grant, Bearer auth)
  • payload.py — structured art-item data with [PREDICT] placeholders for RPT-1
  • config/agents.yaml, config/tasks.yaml — YAML definitions (method names must match decorator names)

Evidence documents (plain text, loaded into the Grounding Service pipeline) are in exercises/data/documents/.

Build & Run

# Activate virtual env (Windows)
.\env\Scripts\Activate.ps1

# Install dependencies (only needed once)
pip install litellm crewai python-dotenv

# Run the crew
python main.py

Requires a .env file in project/Python/starter-project/ (copy structure from the exercise docs):

AICORE_CLIENT_ID=
AICORE_CLIENT_SECRET=
AICORE_AUTH_URL=
RPT1_DEPLOYMENT_URL=
AICORE_RESOURCE_GROUP=
AICORE_BASE_URL=

Conventions

  • CrewAI YAML config: agent/task names in agents.yaml / tasks.yaml must exactly match the Python method names decorated with @agent / @task.
  • Tool pattern: tools are plain functions decorated with @tool("Descriptive Name"). Return error messages as strings so the LLM can handle failures gracefully.
  • LLM model strings: use sap/<model-name> format (e.g. sap/gpt-4o) matching deployments in SAP AI Launchpad.
  • Process: always Process.sequential — tasks pass outputs as context to the next task in order.
  • RPT-1 payload: [PREDICT] string is the placeholder for values to be inferred; schema (dtype, categories, value ranges) must be exact.

Pitfalls

  • Hardcoded Grounding pipeline ID in call_grounding_service() — replace with your own vector DB pipeline ID from SAP AI Launchpad before running.
  • Token refresh: RPT1Client fetches the OAuth token once at init; long-running crews may hit expiry — re-instantiate if needed.
  • No .env validation at startup — credential errors only surface on the first API call.
  • YAML / decorator name mismatch causes silent CrewAI failures with no clear error message.
  • Grounding pipeline must be pre-loaded with the evidence documents; empty pipelines return no results and agents will hallucinate.