iamzulx/crypto-rag

What it claims

All market data is fetched without API keys from public sources: spot prices from six exchanges (Binance, OKX, Bybit, KuCoin, Kraken, Coinbase), Binance Futures, DefiLlama, Mempool.space and the Fear & Greed Index. Only the optional LLM layer needs a key. The README says live numbers "never pass through the LLM", so prices cannot be hallucinated.

Beyond what the summary covers, the tool also offers:

• Cross-exchange spread and arbitrage comparison
• Order-book slippage estimates for a given notional amount
• OHLC trend analysis with an SMA
• BTC mempool, fee and halving data, Ethereum EIP-1559 gas, and Solana inflation
• A watch mode with conditional price alerts
• A portfolio tracker with profit and loss
• Chat follow-up memory, so "berapa harganya sekarang" resolves to the previously named coin
• A coin-category filter over 615 tagged coins

How to use it

Install and build the indexes once:

python3 -m venv venv
venv/bin/pip install requests numpy faiss-cpu fastembed websocket-client
venv/bin/python fetch_data.py
venv/bin/python index.py
venv/bin/python knowledge_index.py

Optionally, enrich coin categories from CoinGecko and then rebuild:

venv/bin/python enrich_categories.py --pages 1
venv/bin/python index.py

To enable the LLM layer, copy config.example.json to config.json and set OPENAI_BASE_URL, OPENAI_API_KEY and OPENAI_MODEL. The example uses StepFun step-3.5-flash. Environment variables of the same names override config.json. With neither set, the tool returns extractive answers.

Examples:

venv/bin/python rag.py                              # interactive chat (default)
venv/bin/python rag.py "jelaskan impermanent loss"
venv/bin/python rag.py --watch "ETH<1800" SOL
venv/bin/python rag.py "slippage beli BTC 5 juta"
venv/bin/python rag.py "portfolio tambah 0.5 BTC harga 60000"

Useful flags:

• --rule forces the extractive path with no LLM.
• --rerank turns on the cross-encoder.
• --agent forces the tool-calling agent.
• --k N sets the number of retrieved documents (default 5).
• --stream prints the answer token by token.
• --live adds sub-second WebSocket prices.

Retrieval evaluation runs against an 87-query golden set:

venv/bin/python eval_rag.py

The README does not publish any scores from this evaluation.

Performance as stated

• Startup: a one-shot run takes about 15–17 s, mostly spent loading the embedding model. In chat mode the model loads once, and later queries take 0.1–1 s.
• LLM synthesis: adds 8–9 s of API latency. --stream reduces the wait before output appears, and --rule returns instantly.
• Cache TTLs: price 3 s, funding and overview 30–60 s, TVL 5 min. Repeat queries for the same coin are near-instant.
• Non-semantic queries (price, funding, TVL, category, market cap) skip the LLM and are fast once the model is warm.

Limits

• Both the corpus and the example queries are in Indonesian. Intent routing appears to be keyed to Indonesian phrasing, so check before using it with English queries.
• Live data depends on exchange availability. Coins without a trade stream, such as delisted ones, fall back to REST.
• Coin metadata and categories are static snapshots taken at index-build time. Keeping them current means re-running fetch_data.py and index.py.
• The README gives no retrieval-quality numbers, no licence and no test coverage.
• The README describes all output as educational only and not financial advice.