How wtfmadrid/tcs-project works
SupportDesk AI is a read-only support investigation assistant: an executive asks a question in the chat UI, the backend runs a LangGraph workflow, and specialist tools retrieve customer records or policy passages for an evidence-backed response. The product also supports policy PDF upload, indexing, and document listing, plus reproducible synthetic customer data. The README and excerpts support the main runtime links; several workflow details and storage internals are documented but not fully excerpted.
Main entry points: ui.py Chat and uploads · mcp_server.py MCP tool server · workflow.py Question router · policy_store.py PDF indexing and search
Components 18
Chat experience
- Chat and uploads
ui.py - Async backend
app_backend.py
Query orchestration
- Question router
workflow.py - Customer specialist
workflow.py - Policy specialist
workflow.py - Answer synthesis
workflow.py - MCP tool server
mcp_server.py
Customer records
- Customer retrieval tools
mcp_server.py - SQLite retrieval
database.py - Customer database
- Synthetic data generator
seed_db.py
Policy knowledge
- Policy tools
mcp_server.py - PDF indexing and search
policy_store.py - Policy vector store
- Policy PDFs
policies
AI services
- OpenAI model
- Local embeddings
policy_store.py
Other components
- Support executive
Connections 26
- Support executive → Chat and uploads: asks question
README.md - Chat and uploads → Async backend: sends request
README.md - Async backend → MCP tool server: opens MCP session
app_backend.py - Async backend → Question router: runs workflow
app_backend.py - Question router → Customer specialist: routes customer
README.md - Question router → Policy specialist: routes policy
README.md - Customer specialist → MCP tool server: uses customer tools
README.md - Policy specialist → MCP tool server: uses policy tools
README.md - Customer specialist → Answer synthesis: returns evidence
README.md - Policy specialist → Answer synthesis: returns evidence
README.md - Answer synthesis → OpenAI model: synthesizes answer
workflow.py - Question router → OpenAI model: classifies question
workflow.py - Customer specialist → OpenAI model: generates response
- Policy specialist → OpenAI model: generates response
- Answer synthesis → Chat and uploads: returns result
README.md - MCP tool server → Customer retrieval tools: registers tools
mcp_server.py - Customer retrieval tools → SQLite retrieval: calls retrieval
mcp_server.py - SQLite retrieval → Customer database: reads records
database.py - Synthetic data generator → Customer database: seeds records
README.md - MCP tool server → Policy tools: registers tools
mcp_server.py - Policy tools → PDF indexing and search: calls search and ingest
mcp_server.py - PDF indexing and search → Policy vector store: stores and searches
policy_store.py - Chat and uploads → Async backend: requests indexing
ui.py - PDF indexing and search → Local embeddings: embeds text
policy_store.py - Chat and uploads → Policy PDFs: saves uploaded PDF
ui.py - PDF indexing and search → Policy PDFs: extracts pages
policy_store.py