AI-First CRM (HCP Module) — Log Interaction Screen

Tech stack (per task):

• Frontend: React + Redux (Toolkit), Google Inter
• Backend: Python FastAPI
• AI Agent: LangGraph
• LLM: Groq (gemma2-9b-it)
• DB: Postgres (Docker Compose)

Prereqs

• Node.js 18+
• Python 3.10+
• Docker Desktop

Run (end-to-end)

1) Start Postgres

docker compose up -d

2) Backend (FastAPI + LangGraph)

cd backend
python -m venv .venv
.\.venv\Scripts\pip install -r requirements.txt
copy .env.example .env

Set GROQ_API_KEY in backend/.env.

Run the API (dev mode with reload):

.\.venv\Scripts\uvicorn app.main:app --reload --port 8000

Seed sample HCPs:

Invoke-RestMethod -Method Post -Uri http://127.0.0.1:8000/api/seed

3) Frontend (React + Redux)

cd frontend
npm install
npm run dev

Open http://localhost:5173/.

What’s implemented

• Structured form matching image.png (Log HCP Interaction).
• Conversational chat ("AI Assistant") that calls the backend LangGraph agent.
• Chat → Form autofill:
  • Example input in the AI panel:
    Today I met with Dr. Sharma and discussed Product X efficacy, the sentiment was positive and I shared the brochures.
  • The system automatically:
    • selects HCP Name (e.g. Dr. Sharma) when a match exists in the HCP list,
    • sets Sentiment (Positive / Neutral / Negative),
    • fills Topics Discussed with the clinical/business discussion only,
    • fills Materials Shared (e.g. "Brochure shared"),
    • highlights updated fields on the form so you can review and manually fix anything.
• Log button → DB:
  • After reviewing/editing the form, press Log to create an interaction row in Postgres via POST /api/interactions.
  • A small status message in the AI panel confirms success or shows validation errors.

LangGraph tools & backend normalization

• LangGraph tools (minimum 5) defined in backend/app/agent.py:
  • tool_log_interaction (LLM summarization/entity extraction + structured patch)
  • tool_edit_interaction (edit flow, currently routed to same logic)
  • tool_search_hcp (placeholder)
  • tool_fetch_product_info (placeholder)
  • tool_suggest_followups (placeholder)

The backend agent normalizes LLM responses so /api/agent/chat always returns a structured suggested_patch with this field set: hcp_name, interaction_type, attendees, topics_discussed, outcomes, follow_up_actions, materials_shared, samples_distributed, sentiment
— even if the raw model output uses different key names. This keeps the chat-driven autofill robust and predictable.