RetailIQ

Retail and supply chain intelligence that runs on your own machine. You upload CSVs or PDFs, ask questions in plain English, and get answers back — no external APIs, no data leaving your machine.

Built to work with real inventory files, sales exports, and supplier docs without pushing anything to a cloud service. Everything runs locally through Ollama.

What it does

Inventory Q&A — drop in a product catalog or inventory CSV and ask things like "which items are below reorder level" or "what's our stock situation for electronics." Uses ChromaDB + LangChain RAG under the hood.

Demand forecasting — trains a small LSTM on your sales history and predicts next-period demand per product. MLflow tracks each training run so you can compare results over time. Works best with 12+ months of data per product.

Supply chain risk — a ReAct agent that searches through uploaded supplier contracts and inventory data. Good for questions like "which suppliers have the longest lead times" or "any contracts expiring soon."

Customer sentiment — same RAG pipeline but against review data. Useful for "what are customers complaining about for product X" type queries.

Data monitoring — Evidently AI for quality checks and drift detection. Upload a CSV, get back a report on missing values and distribution shifts.

Requirements

• Python 3.11
• Ollama installed
• Docker (optional)

Getting started

# first time only
bash setup.sh

# start everything
bash start.sh

Pulls llama3, creates the venv, installs deps, and starts Ollama + MLflow + FastAPI + Streamlit. Open http://localhost:8501.

The sample data in data/sample/ loads automatically on backend startup so all five tabs have something to show right away.

Docker

docker compose build
docker compose up -d

Same ports, same behavior. Ollama runs in its own container and pulls llama3 on first boot (~4.7 GB, takes a few minutes). Subsequent starts use the cached volume.

Running services manually

# terminal 1
ollama serve

# terminal 2
source .venv/bin/activate
mlflow server --host 0.0.0.0 --port 5001 --backend-store-uri ./mlruns

# terminal 3
source .venv/bin/activate
PYTHONPATH=$(pwd) uvicorn backend.api.main:app --host 0.0.0.0 --port 8000 --reload --reload-dir backend --reload-dir frontend

# terminal 4
source .venv/bin/activate
PYTHONPATH=$(pwd) streamlit run frontend/app.py --server.port 8501

Streamlit at :8501, FastAPI docs at :8000/docs, MLflow at :5001, Ollama at :11434.

Environment variables

Copy .env.example to .env — defaults work for local setup out of the box.

Variable	Default
OLLAMA_BASE_URL	http://localhost:11434
OLLAMA_MODEL	llama3
CHROMA_PERSIST_DIR	./data/chroma_db
MLFLOW_TRACKING_URI	http://localhost:5001
EMBEDDING_MODEL	sentence-transformers/all-MiniLM-L6-v2

If you change OLLAMA_MODEL, run ollama pull <model> first.

Sample data

Loaded automatically on startup:

• inventory.csv — 15 products with stock levels, reorder points, supplier IDs
• sales_history.csv — monthly units sold per product across 2023
• supplier_info.csv — 5 suppliers with lead times, reliability scores, risk levels
• customer_reviews.csv — 16 reviews across 5 products with ratings
• retail_qa.json — QA pairs for LoRA fine-tuning

To re-seed manually: PYTHONPATH=$(pwd) python scripts/seed_data.py

Fine-tuning

source .venv/bin/activate
PYTHONPATH=$(pwd) python backend/finetuning/lora_finetune.py data/sample/retail_qa.json

Saves the fine-tuned model to data/peft_model/. Uses DistilBERT + LoRA by default.

License

MIT