An AI-powered analytics assistant that lets business users ask questions about Order Appetit’s MongoDB data in plain English and get precise, tabular insights — no SQL or BI tool required.

This prototype evolved into a production-grade solution. It demonstrates end‑to‑end problem solving: multi‑agent orchestration, schema understanding, query generation, execution, and presentation in a clean Streamlit UX.

• Conversationally answers questions about sales, products, and restaurants using MongoDB data.
• Classifies intent and routes between general chat and task‑specific data analysis.
• Analyzes schemas, resolves naming inconsistencies (e.g., “Mac n Cheese” vs “Mac & Cheese”), and builds optimized MongoDB pipelines.
• Executes queries safely, then formats results into readable, granular tables.
• Persists conversation context with short‑term memory for better follow‑ups.

• Product thinking: focuses on business KPIs/end‑users, not just LLM demos.
• Solid architecture: CrewAI multi‑agent system + LangChain tools + Streamlit UX.
• Practical retrieval: OpenAI embeddings + Pinecone index for item name normalization.
• Data rigor: schema‑aware analysis, query optimization, JSON‑first outputs.
• Shipping mindset: Dockerized, .env‑driven config, clean modular code, clear roadmap.

• Natural‑language analytics on MongoDB collections
• Multi‑agent CrewAI pipeline (Schema Analyzer, Query Builder, Data Analyst)
• Local schema loading for low‑latency analysis and consistency
• Item name normalization via embeddings + Pinecone lookup
• Streamlit chat UI with conversation threads and lightweight memory

• Backend: Python 3.12, CrewAI, LangChain, Pydantic
• LLMs/Embeddings: OpenAI (Chat + Embeddings)
• Vector DB: Pinecone
• Data: MongoDB (via PyMongo), ZenML pipelines
• UI: Streamlit
• Containerization: Docker, docker‑compose

• Conversational router: classifies queries as General vs Task‑Specific.
• CrewAI agents:
  • Schema Analyzer: loads schemas, identifies collections/fields, handles naming variance.
  • Query Builder: generates optimized aggregation pipelines and Python code to execute.
  • Data Analyst: validates/executes code, returns structured tables only.
• Tools: MongoDB schema analyzer, local schema reader, Python REPL executor, Pinecone‑backed item matcher.
• Memory: sliding window buffer for concise, useful context carry‑over.

• Clone: git clone https://github.com/iqbal-sk/Order-Appetit.git && cd Order-Appetit
• Python: 3.12 recommended (Docker path below is easiest)

Create a .env file at repo root or export via your shell. Minimum required:

• OPENAI_API_KEY: OpenAI API key
• PINECONE_API_KEY: Pinecone API key (for item equivalence)
• mongodb_uri: MongoDB URI
• database_name: Target database name
• Optional LangSmith/telemetry: LANGCHAIN_API_KEY, LANGCHAIN_PROJECT, LANGCHAIN_ENDPOINT

• Build + start: docker compose up --build
• App: open http://localhost:8501

Notes:

• The container runs Streamlit from dashboard/src/dashboard/conversational_chatbot.py.
• The code also contains an alternate agent setup in chatbot.py; the Flow‑based version is default.

• Install deps: pip install -r requirements.txt && pip install crewai~=0.76.9 crewai-tools~=0.13.4
• Start UI: cd dashboard/src/dashboard && streamlit run conversational_chatbot.py

• “Give sales of biryani till now.”
• “Which month has the highest sales of pasta?”
• “Top 10 restaurants by sales in the last 12 months.”
• “What are the sales of mac n cheese for the last 7 months?”

• Agents/tasks: dashboard/src/dashboard/config/agents.yaml, dashboard/src/dashboard/config/tasks.yaml
• Schemas mapping: dashboard/src/dashboard/config/schema.yaml
• Local schema JSONs: dashboard/src/dashboard/schemas/*.json

You can tailor agent goals/instructions for your data domain and tighten/relax the Data Analyst’s behavior to be “table‑only, no commentary”.

• Product normalization uses embeddings to find semantically similar item names.
• ZenML pipelines:
  • product_data_pipeline: fetch → clean/standardize products from MongoDB.
  • product_embedding_pipeline: generate OpenAI embeddings → upsert into Pinecone.
• Steps live in dashboard/src/dashboard/steps/* and dashboard/src/dashboard/pipelines/*.

Run the embedding pipeline after setting env vars to populate the Pinecone index used by the chatbot’s item‑matching tool.

dashboard/
  src/dashboard/
    callbacks/           # Agent/Task progress to UI
    config/              # Agents, tasks, schema configs
    memory/              # ConversationBufferWindow
    pipelines/           # ZenML pipelines
    schemas/             # Local schema JSONs
    steps/               # ZenML steps (Mongo, embeddings, Pinecone)
    tools/               # Mongo analyzer, item finder, Python executor
    ui/                  # Streamlit chat + sidebar + CSS
    utils/               # Crew assembly, summarizers, chat state helpers
    chatbot.py           # Alt crew setup (legacy)
    conversational_chatbot.py  # Flow + router entrypoint

• Limitations:
  • No charts/visualizations yet (tables only)
  • Follow‑up query understanding can be improved further
  • Some latency on first‑run model/tools initialization
• Roadmap:
  • Add plotly/altair‑based charts and CSV export
  • Query caching and results persistence
  • Role‑based access control + audit logs
  • Observability (LangSmith) and eval harness
  • Expand semantic normalization beyond products (e.g., restaurant names, categories)

• Keep secrets in .env or your secrets manager; never commit keys.
• Pipelines read‑only where possible; writing to Pinecone is scoped to the configured index.