How muthukumar-selvarasu/agentic-customer-support works
This repository implements an online-shop support desk, with terminal and web interfaces that expose each turn’s decisions. The principal workflow accepts a customer message, applies security checks, consults the support agent and shop tools, masks sensitive reply data, saves customer memory, and records a trace. The README also identifies MCP tools, A2A security services, and evaluation as product capabilities. Several pipeline sections and the web implementation are unsampled, so edges that depend on their internal calls are omitted or marked unverified.
Main entry points: pipeline.py Turn pipeline · web.py Web app · run.py Evaluation runner · cli.py Terminal app
Components 18
Customer interfaces
- Terminal app
support/cli.py - Web app
support/web.py
Turn orchestration
- Turn pipeline
support/pipeline.py - Support agent
support/pipeline.py
Security services
- Input sanitizer
guards/sanitizer.py - Security Judge
guards/judge/app.py - Attack patterns
guards/judge/patterns.py - Topic guardrail
guards/guardrail.py - Data Masker
guards/masker/app.py - Masking logic
guards/masker/mask.py
Shop data and tools
- MCP shop tools
mcp_toolbox/tools.yaml - Shop database
db/seed.sql
Memory and evaluation
- Customer memory
support/memory.py - Trace telemetry
support/telemetry.py - Evaluation runner
eval/run.py - Evaluation cases
eval/gold.py - Evaluation scoring
eval/score.py
Other components
- Customer
Connections 22
- Customer → Terminal app: sends message
README.md - Customer → Web app: sends message
README.md - Terminal app → Turn pipeline: runs turn
support/cli.py - Web app → Turn pipeline: runs turn
support/web.py - Turn pipeline → Input sanitizer: checks input
README.md - Turn pipeline → Security Judge: requests verdict
support/pipeline.py - Security Judge → Attack patterns: scans message
guards/judge/app.py - Turn pipeline → Topic guardrail: checks topic
support/pipeline.py - Turn pipeline → Support agent: dispatches turn
README.md - Support agent → MCP shop tools: calls tools
README.md - MCP shop tools → Shop database: reads or updates
README.md - Turn pipeline → Data Masker: requests masking
README.md - Data Masker → Masking logic: masks reply
guards/masker/app.py - Turn pipeline → Customer memory: recalls and saves
README.md - Turn pipeline → Trace telemetry: records trace
support/pipeline.py - Terminal app → Trace telemetry: sets up tracing
support/cli.py - Web app → Trace telemetry: uses tracing
README.md - Evaluation runner → Web app: drives scenarios
eval/run.py - Evaluation runner → Evaluation cases: loads cases
eval/run.py - Evaluation runner → Evaluation scoring: scores results
eval/run.py - Evaluation scoring → Evaluation cases: uses expected outcomes
eval/score.py - Evaluation scoring → Customer memory: checks memory
eval/score.py