Research Copilot Pro

An Indian-market research terminal with a TradingAgents-style multi-agent desk. Live NSE/BSE quotes, index coverage, a rule-based quant screen, and a full analyst → bull/bear → risk → portfolio-manager debate that ends in a rated, sized trade call.

Educational research tool. Not investment advice. LLM output can be wrong — verify every level before trading. See Data & legal.

---

What it does

• Live multi-source quotes. NSE website JSON first, Yahoo 1-minute API second, yfinance last. Every price shows its source, the exchange timestamp, and a stale flag if the market is open but the tick is old. Auto-refreshes every few seconds.
• Index coverage. Nifty 50, Bank Nifty, Sensex, Nifty IT and India VIX (plus USD/INR and Brent) in a live strip — and searchable like any stock, with full price, chart and quant detail.
• Honest watchlist. Symbols join the watchlist only when you explicitly save them (★ button). Searching never pollutes it.
• Rule-based quant screen. EMA stack, RSI-14, MACD histogram, ATR and factor scores (Value / Quality / Growth / Momentum / Volatility) computed in Python — no LLM in the loop.
• Multi-agent AI research. A TradingAgents-style pipeline: four analysts in parallel → Bull vs Bear debate → Research Manager → Trader → Aggressive/Conservative/Neutral risk debate → Portfolio Manager, producing a 5-tier rating with entry / stop / target and a position-size check.
• Screener.in-style fundamentals. The Fundamentals analyst reads a data pack that layers computed ratios (net debt, D/E, interest coverage, current ratio, working capital, ROCE/ROE, OCF-vs-PAT cash conversion, YoY/QoQ growth, cash conversion cycle) plus annual and quarterly income statement, balance sheet and cash flow on top of Yahoo's headline multiples — all in ₹ crore, all computed by code so the model only cites verified figures.
• Decision log & reflection. After the holding window each call is scored against Nifty/Sensex, reflected on, and the lessons feed later Portfolio Manager prompts.
• Provenance. Every number is computed by code; the LLM only interprets. Each run records which data source answered what.

The multi-agent pipeline

   Analysts (parallel)          Research                Trading        Risk (debate)        Portfolio
 ┌─────────────────────┐   ┌───────────────────┐   ┌───────────┐   ┌──────────────────┐   ┌──────────────┐
 │ Market   Sentiment  │   │  Bull  ⇄  Bear    │   │           │   │ Aggressive       │   │              │
 │ News     Fundamentals├──▶│        ↓         ├──▶│  Trader   ├──▶│ Conservative     ├──▶│ Portfolio Mgr│
 └─────────────────────┘   │  Research Manager │   │           │   │ Neutral          │   │  (final call)│
                           └───────────────────┘   └───────────┘   └──────────────────┘   └──────────────┘

Depth controls debate rounds (Shallow 1 / Medium 3 / Deep 5). Output language includes Hinglish.

Screenshots

Overview — index with chart & quant screen	AI Research — the multi-agent desk

Index search returns full detail for benchmarks too (here, Sensex on BSE):

Quick start

pip install -r requirements.txt
copy .env.example .env        # on Mac/Linux: cp .env.example .env
# paste ONE LLM API key into .env
uvicorn server:app --port 8000

Open http://localhost:8000.

Offline tests (no network, no keys):

pip install pytest httpx
python -m pytest tests -q

Configuration

Pick any one provider in .env (or per-run in the UI): Anthropic, OpenAI, Google, DeepSeek, Groq, OpenRouter, Ollama (local), or any OpenAI-compatible endpoint. Deep-thinking models drive the manager / trader / portfolio manager; quick-thinking models drive analysts and debaters.

Env var	Purpose
ANTHROPIC_API_KEY, OPENAI_API_KEY, GROQ_API_KEY, …	one key for your chosen provider
COPILOT_LLM_PROVIDER	force a provider
COPILOT_DEEP_THINK_LLM / COPILOT_QUICK_THINK_LLM	model ids
COPILOT_OUTPUT_LANGUAGE, COPILOT_MAX_DEBATE_ROUNDS, COPILOT_RISK_PER_TRADE_PCT, …	run defaults

Reports

Completed runs write a full report tree under runs/<id>/ and a combined complete_report.md. A real end-to-end sample is published at reports/UNIPARTS.md; more (e.g. POLICYBZR) will follow — see reports/README.md.

Provider request-size limits

Small/free LLM tiers cap tokens per request (e.g. Groq on-demand ≈ 8 000 TPM), and a naive pipeline easily sends a 9 000+ token debate prompt, dying with HTTP 413. This project keeps every call under the limit by capping what gets embedded: each analyst report is truncated to report_char_budget chars and accumulated debate history to debate_history_chars chars (both tunable via COPILOT_REPORT_CHAR_BUDGET / COPILOT_DEBATE_HISTORY_CHARS). Debater replies are also token-bounded so later turns stay small.

If a price looks wrong

Open http://localhost:8000/api/diag/RELIANCE?ex=NSE. It lists what each source returned for the symbol, with timings and the spread between them, so a bad tick is traceable to its source.

Project layout

server.py            FastAPI app: quotes, indices, search, chart, runs, reports
copilot/
  market.py          live NSE/Yahoo/yfinance quotes, indices, history, search
  dataflows.py       data packs per analyst (fundamentals, news, sentiment…)
  agents.py          the analyst / debater / manager / trader / PM prompts
  graph.py           orchestrates the pipeline
  indicators.py      EMA / RSI / MACD / ATR (pure Python)
  jobs.py            run manager, decision log, reflection
  reporting.py       markdown report tree
web/                 the dashboard (vanilla JS, no build step)
tests/               offline test-suite

Data sources & legal

NSE and Yahoo endpoints are unofficial, personal-use only and can change or block requests at any time — use a licensed feed for any public or paid product. Publishing buy/sell calls publicly may require SEBI Research Analyst registration. This project is educational and is not investment advice.

License

MIT

---

Built and maintained by Khushal Jain