ROIC Chart Genie

What it does

ROIC Chart Genie lets you type "Compare Apple's ROIC to Microsoft's debt-to-equity over the last ten years" and get back a PNG chart—no tickers, no Excel, zero manual scraping.

GPT-3 reads your prompt and extracts one- or two-stock symbols.
A mini-crawler pulls 24 key financial metrics from ROIC.ai and saves each company to a CSV.
GPT writes the Matplotlib code needed to graph exactly what you asked.
The server executes the code headlessly, saves the plot, encodes it in base-64, and returns JSON (prompt, image, code).

Everything runs behind a single Flask endpoint: POST /prompt.

Pipeline at a glance

flowchart LR
  U(User prompt) -->|GPT symbols| S[get_symbols()]
  S --> CSV[request_stock() → CSV]
  U -->|CSV list + prompt| G[get_graph()]
  G --> C[execute_code()]
  C --> IMG[.png → base64]
  IMG --> R(Return JSON)

Key pieces

Module / file

Job description

get_symbols (GPT)

Pull comma-separated tickers from messy prose

request_stock

Scrape ROIC.ai → clean → DataFrame → TICKER.csv

get_graph (GPT)

Inject file list + prompt into a code-completion template

execute_code

Run the generated Python, swap plt.show() for plt.savefig()

Flask route

Orchestrates flow, catches errors, sends image back to browser

Try it locally

pip install flask flask-cors openai requests beautifulsoup4 lxml pandas matplotlib
export OPEN_AI_KEY="sk-your-key"
python app.py

Then:

curl -X POST http://localhost:5000/prompt \
     -d "plot google vs amazon revenue per share 2010-2022"

The response JSON holds a Base-64 PNG ready for <img src="data:image/png;base64, …">.

Tweak knobs

Variable

Purpose

configs['stocks']

Prompt prefix that tells GPT how to spot tickers

categories

24 scraped metrics; trim to speed things up

max_tokens in GPT calls

Trade creativity for cost / determinism

execute_code()

Swap Matplotlib backend or add seaborn styling

Caveats

Scraper fragility – ROIC.ai class names may change without warning.
GPT misfires – Irrelevant words like "CAPEX" sometimes sneak in; we strip a few common offenders.
Two-stock limit – Code template handles one or two CSVs; extend it for baskets.
Security – Generated Python is executed with exec; sandbox before deploying in the wild.

Next steps

Cache scraped CSVs in Firestore to avoid repeat hits.
Move generation to gpt-3.5-turbo for cheaper completions.
Add a GET /image/<id> endpoint for CDN-friendly serving.
Let GPT select chart types: bars for margins, lines for EPS, etc.

Paste a question, get a chart—no finance terminal required.