Finance RAG — Time‑Series Insights

A Retrieval‑Augmented Generation (RAG) app that combines market data (yfinance), vector search (ChromaDB), and interactive charts (Streamlit + Matplotlib) to surface daily insights with temporal context (SMA, 1‑day, 1‑month deltas).

Live stack: Streamlit • ChromaDB • SentenceTransformers • yfinance • Matplotlib

---

What this app does

• Embeds daily market facts (close/open/high/low/volume) and stores them in a Chroma vector DB.
• Retrieves facts via semantic search with optional ticker filter.
• Analyzes time‑series (1‑month window): last close, 1‑day Δ, 1‑month Δ, and SMA5/SMA10 overlays.
• Ready for Streamlit Cloud deployment.

---

Repository structure

fin_rag_project/
├─ app.py                  # Streamlit app (entrypoint)
├─ vector_store/           # Chroma persistent store (created locally)
├─ requirements.txt        # Python deps for Streamlit Cloud
├─ .gitignore
└─ README.md               # This file

If you used a Notebook to ingest data, you can later move that code into an ingest.py script. For now it’s fine that you ingested locally.

---

How it works (high level)

1. Ingestion (done locally):

   • Pull last month of OHLCV data from yfinance for tickers (e.g., TSLA, GOOGL, MSFT, TECHM.NS, INFY.NS, TCS.NS, RELIANCE.NS).
   • Normalize data (handles MultiIndex, ensures Date column, uses Close/Adj Close).
   • Create daily “facts” strings and embed with SentenceTransformers/all‑MiniLM‑L6‑v2.
   • Store in Chroma persistent collection: finance-docs.

2. Query (Streamlit app – app.py):

   • User enters a natural language question and optional ticker.
   • We do a vector search in Chroma and show the top‑k matches.
   • We also fetch 1‑month of latest prices and plot SMA5/SMA10 + summary metrics.

---

Local development (Windows / Anaconda)

# 1) Create & activate env (if you don’t already have one)
conda create -n finrag python=3.11 -y
conda activate finrag

# 2) Install deps
pip install -r requirements.txt

# 3) Run the app
streamlit run app.py

The app expects a local Chroma collection at:
vector_store/ with collection name finance-docs.

If the collection is empty (first run on a new machine), app.py has a tiny seed so queries don’t crash on Cloud (see “Seeding” below).

---

Deploy to Streamlit Cloud

1. Push this repo to GitHub (you already did).
2. Go to share.streamlit.io → Create app.
3. Select your repo: sripranay/fin_rag_project
4. Branch: main
5. App file path: app.py
6. Hit Deploy.

Streamlit Cloud will install from requirements.txt and run app.py.

---

Requirements

Minimal requirements.txt

streamlit>=1.36
chromadb>=0.5
sentence-transformers>=2.7
yfinance>=0.2.40
pandas>=2.2
numpy>=1.26
matplotlib>=3.8

---

Seeding the vector store (for first-time Cloud runs)

Streamlit Cloud doesn’t have your local vector_store/ yet. To avoid errors on the very first query, app.py contains a tiny seed that runs if the collection is empty:

# After: collection = client.get_collection(COLLECTION)
try:
    existing = collection.peek(1)
except Exception:
    existing = {"documents": []}

if not existing.get("documents"):
    seed_docs = ["Seed fact: RELIANCE.NS is listed on NSE."]
    seed_meta = [{"ticker": "RELIANCE.NS", "date": "seed"}]
    seed_ids  = ["seed-1"]
    from sentence_transformers import SentenceTransformer
    _m = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
    seed_emb = _m.encode(seed_docs, normalize_embeddings=True).tolist()
    collection.add(documents=seed_docs, metadatas=seed_meta, ids=seed_ids, embeddings=seed_emb)

This ensures the app still responds even before you ingest real data into Chroma on the server.

---

Ingestion (what we ran)

You ran ingestion locally (Notebook) with logic equivalent to:

• Download 1‑month daily bars via yfinance
• Normalize columns/dates
• Compose fact strings per date
• Embed with all‑MiniLM‑L6‑v2
• col.add() to the finance-docs collection

You can move this into a script (e.g., ingest.py) later if you want a CLI like: python ingest.py --tickers TSLA,GOOGL,MSFT

---

Query flow (at runtime)

• Encode the user question → query Chroma (top_k).
• Optional { "ticker": "RELIANCE.NS" } filter.
• Show the ranked hits (doc + metadata + distance).
• Render chart & metrics via yfinance live pull (1 month window).

---

Notes on data / compliance

• This app uses public market data via yfinance; it is educational.
• No PII is collected.
• If you add news/filings in the future, verify usage rights and attribution.

---

Troubleshooting

• “No results”: Remove ticker filter or ingest more data.
• Chroma not found: Ensure vector_store/ exists and is readable.
• Cloud errors: Check requirements.txt formatting and app path (app.py).
• Plots not showing: Do not set custom Matplotlib styles on Cloud; the app uses default settings.

---

Roadmap

• Add news and filings chunking/ingestion.
• Add risk metrics (volatility, drawdown).
• Add RAGAS for retrieval evaluation.
• Add period dropdown (1/3/6/12 months).

---

📄 License

MIT (or add your preferred license).

---

Authors

Sri Pranay — project owner
Assistant — setup guidance & docs

---

Quick commands (Git)

# from your repo folder
git add README.md
git commit -m "Add complete README with setup, usage, and deployment"
git push