How Claw Chat works¶

Claw is a chat-style analyst that lives on the Claw page. You ask a free-form question, Claw reads your own signal history, and it answers — citing the specific signals it used.

If you're curious about the retrieval stack under the hood, see Claw chat (RAG). This page is about what Claw actually promises you as a user.

Four rules Claw always follows¶

1. It only reads your signals¶

Claw doesn't search the open web, doesn't check Twitter, doesn't scrape CNBC on the fly. When you ask "What happened with NVDA this week?", it looks through the signals that your pipeline has already captured for your watchlist — the same signals you can see in the dashboard feed.

Why this matters: the answer is grounded in concrete things you can verify. If Claw says "NVDA's CFO stepped down on Tuesday," there is a signal card behind that statement.

2. It cites every claim¶

Every factual sentence in Claw's answer has a bracketed number like [1], [2], [3]. Those numbers key to specific signal cards — you can find them on your dashboard feed and trace back to the original source (news article, SEC filing, YouTube transcript, Reddit thread, etc.).

Why this matters: no more "LLM said something, is it true?" You can audit the basis for every claim in ~2 clicks.

3. It refuses when it doesn't know¶

If your signal history doesn't cover the topic you're asking about, Claw will say so instead of making something up. A typical refusal:

"I don't have any signals in your history that are relevant enough to answer that confidently. Try rephrasing with different keywords, or wait for the next pipeline cycle to bring in fresh signals."

This is deliberate. The alternative — where a language model confidently fabricates a plausible-sounding answer from unrelated context — is the single most common way RAG systems mislead users, and the single most expensive failure mode for anything related to investing.

Why this matters: silence is safer than fiction. If Claw tells you something, it's because real signals back it.

4. It never gives investment advice¶

Claw will not tell you what to buy, what to sell, when to act, or what position to take. Ask it "Should I buy NVDA?" and it will politely decline and redirect you to talk to a licensed professional.

Claw is an information tool, not a financial adviser. It summarises and contextualises market events from your signals — that's it.

Why this matters: compliance, liability, and honesty. Claw is very good at what happened and why. It has no business opinion on what you should do about it.

What signals reach Claw¶

Claw only knows what the pipeline has already crawled on your behalf. The full list of sources — news outlets, SEC filings, subreddits, newswires, central banks, economic-indicator feeds — is in Where your signals come from, along with an honest note on what's not covered.

Questions Claw handles well¶

"Why did NVDA drop 3% on Tuesday?" → pulls recent NVDA signals, synthesises the story
"Summarise this week's FLASH signals" → retrieves high-urgency items, returns a roll-up
"What's the current Chinese semiconductor policy situation?" → aggregates across multiple signals and agents
"最近有哪些关于 AI 芯片需求的信号?" → full Chinese round-trip works, answer comes in the same language you asked in

Questions Claw won't help with¶

Question	Why not	Where to go instead
"What's NVDA trading at right now?"	Claw reads signals, not live prices.	Dashboard index bar or stock card.
"Should I buy / sell / hold X?"	Claw refuses to give advice.	Talk to a licensed adviser.
"What are other users asking about?"	Row Level Security hides other users' data from your session.	There is no cross-user view — by design.
"Make up a bull case for TSLA"	Claw only reasons from real retrieved signals, not imagination.	—

What the answer looks like¶

Short. Usually under 250 words, unless your question genuinely needs a list.
Same language as your question. Ask in English → English answer. Ask in 中文 → 中文 answer.
Cited. [1] [2] numbers link to the signal cards used.
Session-persistent. Your past conversations are kept in the browser (localStorage) and survive page reloads. Every question and answer is also written to a server-side query log for audit.