🎤 Workshop context: Built hands-on at the AI Builder's Mixer: Delhi Edition

(Microsoft Fabric User Group: Uttar Pradesh Power BI Club), held on 5th July 2026 at Masters' Union, CDS Corporate Tower, New Delhi.

Original architecture, problem statement, and starter repo by Divij Bajaj (@divijbajaj7) — the session's speaker. I built this live during the workshop, implementing the multi-agent router/specialist/supervisor pattern from his problem statement myself.

What I did: Implemented and tested the LLM-powered Router Agent's query classification across all 6 routes (INVENTORY / PRICING / REVIEWS / CATALOG / MULTI / GENERAL), wired up the FastAPI backend (/chat, /briefing, /trace endpoints), verified the Supervisor's cross-domain synthesis on MULTI-route queries, and confirmed session-based memory correctly resolved multi-turn follow-ups (e.g. "what's its margin?").

---

RetailMind v2 — Multi-Agent Product Intelligence

A 5-agent LangChain system (Router + 4 Specialists + Supervisor) wrapped in a FastAPI backend with a custom HTML/CSS/JS frontend (landing page + floating chat widget — no Streamlit).

Built for the StyleCraft demo: replaces Priya Mehta's 4–5-hour weekly catalog audit with a real-time conversational interface.

---

Quickstart

cd retailmind_agent
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env          # add OPENAI_API_KEY
python run.py                 # opens http://127.0.0.1:8000 in your browser

Click the chat icon in the bottom-right of the landing page. The Daily Briefing renders on first open (Supervisor orchestrates all four specialists).

---

Architecture

                       [User Query]
                            |
                    +-------v--------+
                    |  Router Agent  |   LLM classifier:
                    |  (gpt-4o-mini) |   INVENTORY / PRICING /
                    +-------+--------+   REVIEWS / CATALOG /
                            |            MULTI / GENERAL
          +-----------------+------------------+
          |        |        |        |         |
          v        v        v        v         v
    +-----------+ +---------+ +---------+ +---------+
    | Inventory | | Pricing | | Reviews | | Catalog |
    |   Agent   | |  Agent  | |  Agent  | |  Agent  |
    +-----+-----+ +----+----+ +----+----+ +----+----+
          |            |           |           |
       (own tools)  (own tools)  (own tools)  (own tools)
          |            |           |           |
          +------------+-----+-----+-----------+
                             |
                             v  (only for MULTI route +
                +------------+------------+   Daily Briefing)
                |  Supervisor / Synthesis |
                |  Agent (gpt-4o, no tools)|
                +------------+------------+
                             |
                             v
                       [Final Answer]

Query → agent path mapping

Route	Triggered when the user asks about…	Dispatched to
INVENTORY	Stock levels, restock needs, days to stockout	InventoryAgent → get_inventory_health / generate_restock_alert
PRICING	Margin, profitability, price positioning	PricingAgent → get_pricing_analysis / compare_category_pricing
REVIEWS	Customer feedback, sentiment, complaints	ReviewsAgent → get_review_insights / get_negative_review_themes
CATALOG	Product search, category overview	CatalogAgent → search_products / get_category_performance
MULTI	Cross-domain ("Why is X underperforming?", "Should we discount Y?")	Supervisor → calls 2+ specialists in parallel, synthesises one prioritised answer with citations
GENERAL	Greetings, meta questions	Direct LLM, no tools

The router is LLM-powered (ChatOpenAI + Pydantic structured output) — no keyword/regex matching. See agents/router.py.

Specialist isolation

Each specialist file imports only its own tools module:

• agents/inventory_agent.py ← tools/inventory_tools.py
• agents/pricing_agent.py ← tools/pricing_tools.py
• agents/reviews_agent.py ← tools/reviews_tools.py
• agents/catalog_agent.py ← tools/catalog_tools.py

There is no shared ALL_TOOLS registry. Code inspection will confirm.

Memory

Per-session BufferMemory (in core/memory.py) keyed by a UUID session_id that the frontend persists in localStorage. Every /chat call loads chat_history into the router and the specialist/supervisor so follow-ups like "what's its margin?" resolve to the SKU mentioned earlier. The Clear button wipes the session and re-triggers the Daily Briefing. This implements the +5 bonus criterion (ConversationBufferMemory equivalent).

---

Project layout

retailmind_agent/
├── run.py                       # python run.py → uvicorn + auto-open browser
├── server.py                    # FastAPI: /chat /briefing /summary /clear /trace /categories
├── requirements.txt
├── .env.example
├── data/
│   ├── retailmind_products.csv
│   └── retailmind_reviews.csv
├── agents/
│   ├── router.py                # LLM-powered Router (6 routes)
│   ├── inventory_agent.py       # InventoryAgent (2 tools)
│   ├── pricing_agent.py         # PricingAgent (2 tools)
│   ├── reviews_agent.py         # ReviewsAgent (2 tools)
│   ├── catalog_agent.py         # CatalogAgent (2 tools)
│   └── supervisor.py            # Synthesis + Daily Briefing
├── tools/
│   ├── _data.py                 # Shared CSV loaders (NOT a tool)
│   ├── inventory_tools.py
│   ├── pricing_tools.py
│   ├── reviews_tools.py
│   └── catalog_tools.py
├── core/
│   ├── memory.py                # BufferMemory + SessionStore
│   ├── prompts.py               # 5 distinct system prompts
│   └── trace.py                 # AgentTrace dataclass
└── frontend/
    ├── index.html               # landing page + floating chat widget
    ├── styles.css               # editorial design — Fraunces + Geist + JetBrains Mono
    └── app.js                   # vanilla ES module chat client

---

API surface

Method	Path	Purpose
GET	/	landing page
GET	/summary	catalog snapshot tiles (no LLM)
GET	/categories	category dropdown values
GET	/briefing?session_id=…	cached or freshly generated Daily Briefing
POST	/chat	{session_id, message, category_filter} → {response, route, reason, specialists}
POST	/clear	reset session + return fresh briefing
GET	/trace/{session_id}	last AgentTrace (router decision + per-step timings)

---

Sample queries to test routing

Query	Expected route
"Which dresses are low on stock?"	INVENTORY
"Margin on SC011?"	PRICING
"Negative themes in Tops?"	REVIEWS
"Top performers in Outerwear"	CATALOG
"Why is SC020 underperforming?"	MULTI (≥2 specialists)
"Should we discount the Velvet Party Dress?"	MULTI
"Hi, what can you do?"	GENERAL
"Stockouts in next 10 days"	INVENTORY (threshold=10 extracted by LLM)

The Agent Trace toggle under each assistant message shows which router decision was made and which specialists were called — useful for evaluator verification.