MarketOps Copilot (WIP)

Time-aware, table-literate, verifiable RAG app. Single-command local run via Streamlit + FAISS + Transformers.

Core features (MVP):

• Document ingestion from data/samples (PDF, TXT, MD).
• Vector search (Sentence-Transformers) + context-constrained answers (FLAN-T5).
• Topic tagging (zero-shot) and optional FinBERT sentiment.
• Table extraction from PDFs and Table QA with TAPAS (cell coordinates returned).
• Simple numeric self-check widget (sum/avg over cited table columns).

This is a lightweight, on-device-friendly version. It avoids external APIs by default.

---

Quickstart

1) Clone / unzip and install

# (Option A) Local venv
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python -c "import nltk; import ssl; ssl._create_default_https_context = ssl._create_unverified_context; import nltk; nltk.download('punkt'); nltk.download('stopwords')" || true

First run will download small models (embeddings, FLAN-T5, TAPAS).

2) Run

streamlit run streamlit_app.py

Open the URL shown (default: http://localhost:8501).

3) Docker (optional)

docker build -t marketops-copilot .
docker run --rm -p 8501:8501 -v $PWD/data:/app/data marketops-copilot

4) Project structure

marketops-copilot/
├─ streamlit_app.py
├─ app/
│  ├─ core/
│  │  ├─ rag.py
│  │  ├─ table_qa.py
│  │  ├─ classify.py
│  │  └─ verify.py
│  └─ ingest/
│     └─ pdf_parse.py
├─ data/
│  └─ samples/ # (placeholders)
│     ├─ notice_example.txt
│     ├─ fees_table.pdf            
│     └─ listing_rules_excerpt.md
├─ requirements.txt
├─ Dockerfile
├─ .gitignore
└─ README.md

5) Add your own docs Drop PDFs / TXTs into data/samples and click Ingest in the left sidebar.

---

Notes

• If you have trouble with camelot, use basic PDF text via pdfplumber.
• TAPAS expects tabular data. If your PDFs have embedded tables, extraction quality varies by file.

---