README.md

metadata

title: Market Performance Sentinel (Demo)
emoji: 🧬
colorFrom: indigo
colorTo: purple
sdk: streamlit
sdk_version: 1.39.0
app_file: app.py
pinned: false
license: mit
short_description: Multi-agent LangGraph chatbot for market analytics.

Market Performance Sentinel — Demo

A multi-agent LangGraph chatbot that answers natural-language questions about pharma market performance — built around a supervisor + specialised agents architecture.

This Space runs on 100% fictional data. All company names, product names, and metric values are synthetic and were generated procedurally for demonstration purposes. The fictional company in the demo is NovaPharma and its products (NOVACOR, NOVAGLU, etc.) do not exist.

What this demo showcases

The chatbot answers questions like:

• "What's the market share of the top 3 products in France for the Growth Hormone market?"
• "In LATAM, which country has the biggest QTR-QoQ value change in the Hypothyroid market in 25Q3?"
• "Which cluster is leading the change in APAC for the Anti-EGFR market?"
• "What are the clusters for the Injectable Platform in APAC?"

It is not a Retrieval-Augmented-Generation chatbot — it is a multi-agent system where each agent has a specific responsibility:

              ┌────────────────┐
              │ query_analyzer │   ← LLM classifies intent
              └───────┬────────┘
        ┌─────────────┼───────────────┐
   data_retrieval  parameter_info  out_of_scope
        │             │                │
        ▼             ▼                │
┌────────────────┐ ┌──────────────────┐│
│ data_extraction│ │  data_knowledge  ││  ← LLM-assisted filter parsing,
│   (SQL query)  │ │  (param tables)  ││    fuzzy parameter lookup
└────────┬───────┘ └────────┬─────────┘│
         │  (optional)      │          │
         ▼                  │          │
┌────────────────┐          │          │
│ leading_country│          │          │  ← deterministic country
│  (delta calc)  │          │          │    delta calculation
└────────┬───────┘          │          │
         │                  │          │
         ▼                  ▼          ▼
         ┌──────────────────────────────┐
         │      generate_response       │  ← LLM synthesises final
         │  (LLM synthesis + fallback)  │    business-friendly answer
         └──────────────────────────────┘

Node	Role
query_analyzer	LLM classifies user intent: data_retrieval, parameter_info, both, or out_of_scope.
data_extraction	LLM parses filters (region, period, product, market, …) from the user query, then issues a targeted SQL query — no full-table scan.
data_knowledge	Keyword-scored lookup into the parameter tables (clusters, markets, country-region mapping).
leading_country	Deterministic, no-LLM calculation that reconstructs prior-period deltas and ranks countries by contribution.
generate_response	LLM synthesises a natural-language answer from extracted data + filters + conversation history; falls back to a deterministic CSV-formatted answer if the LLM is unavailable.

Tech stack

• Streamlit for the UI
• LangGraph for the agent graph
• Hugging Face Inference Providers (InferenceClient) for the LLM calls — free serverless tier
• SQLite (single-file demo.sqlite) as the metrics store — pre-seeded with synthetic data
• Python 3.10+

Configuration

The Space needs one secret:

Secret	Where to set it	Notes
HF_TOKEN	Space → Settings → Variables and secrets	A free Hugging Face read token works fine.

Optional environment variables:

Variable	Default	Description
HF_MODEL	meta-llama/Llama-3.3-70B-Instruct	Model used for all LLM calls. Any OpenAI-compatible chat model on HF Inference Providers will work.
HF_PROVIDER	auto	HF inference provider routing (auto, hf-inference, together, …).

Running locally

pip install -r requirements.txt
python seed_data.py            # creates demo.sqlite (one-off; idempotent)
export HF_TOKEN=hf_xxx          # Linux/macOS
# $env:HF_TOKEN = "hf_xxx"      # Windows PowerShell
streamlit run app.py

Disclaimer

This is a demonstration project. The data, products, and company branding are entirely fictional. Numbers were generated procedurally and have no relationship to any real-world pharmaceutical market.

License

MIT