How vishalna/agent_workshop works
This repository is a local-first refund-agent workshop, not a deployed refund service. Its main participant workflow progresses from an ungrounded LLM response through trusted lookups, policy retrieval, deterministic controls, and a guarded tool-using agent; a separate golden-case stage evaluates deterministic decisions. The stages are standalone teaching examples, not one connected runtime pipeline. Ollama and shared JSON/Markdown files provide local model and sample-data dependencies. The map preserves the README’s major capabilities; unsampled files are represented only where their paths are explicit in the tree or README.
Main entry points: 02_tools/run.py Trusted Lookups · 03_rag/run.py Policy Retrieval · 04_guardrails/run.py Guardrail Demo · 05_agent_loop/run.py Agent Tools
Components 14
Workshop Stages
- LLM-only Demo
01_llm_only/run.py
Trusted Context
- Trusted Lookups
02_tools/run.py - Policy Retrieval
03_rag/run.py - Refund Policy
shared_data/refund_policy.md - Order Records
shared_data/orders.json - Refund History
shared_data/refund_history.json
Decision Controls
- Guardrail Demo
04_guardrails/run.py - Agent Loop
05_agent_loop/run.py - Agent Tools
05_agent_loop/run.py
Evaluation
- Golden-case Runner
06_evals/run_evals.py - Decision Logic
06_evals/decision.py - Golden Cases
06_evals/golden_cases.json
Other components
- Participant
- Ollama
Connections 23
- Participant → LLM-only Demo: runs demo
README.md - Participant → Trusted Lookups: runs demo
README.md - Participant → Policy Retrieval: runs demo
README.md - Participant → Guardrail Demo: runs demo
README.md - Participant → Agent Loop: runs demo
README.md - Participant → Golden-case Runner: runs evals
README.md - LLM-only Demo → Ollama: requests generation
01_llm_only/run.py - Trusted Lookups → Order Records: reads
02_tools/run.py - Trusted Lookups → Refund History: reads
02_tools/run.py - Trusted Lookups → Ollama: requests generation
02_tools/run.py - Policy Retrieval → Order Records: reads
03_rag/run.py - Policy Retrieval → Refund Policy: retrieves from
03_rag/run.py - Policy Retrieval → Ollama: embeds and generates
03_rag/run.py - Guardrail Demo → Order Records: reads
04_guardrails/run.py - Guardrail Demo → Refund History: reads
04_guardrails/run.py - Guardrail Demo → Ollama: classifies with
04_guardrails/run.py - Agent Loop → Agent Tools: dispatches tools
05_agent_loop/run.py - Agent Loop → Ollama: requests tool calls
05_agent_loop/run.py - Agent Tools → Order Records: reads
05_agent_loop/run.py - Agent Tools → Refund History: reads
05_agent_loop/run.py - Agent Tools → Refund Policy: searches
05_agent_loop/run.py - Golden-case Runner → Golden Cases: loads cases
06_evals/run_evals.py - Golden-case Runner → Decision Logic: calls
06_evals/run_evals.py