An assistant answers from memory unless it is given somewhere to look, and memory is where invented references come from. Connecting it to a documentation server gives it a place to search and a passage to quote. Artha publishes one over its glossary, explainer wiki and research: one tool, no account, no key, and no access to user data.
Key takeaways
- A model with no source to search will produce a plausible reference rather than admit it has none.
- A documentation server returns titles, links and quoted fragments, so an answer can be checked at its source.
- Public documentation servers need no sign-in, which is why they are separate from servers that reach an account.
- Separation is structural: a server with no path to user data cannot be prompted into reaching one.
- A citation is only useful if you open it; the link is there so the quote can be verified.
Why does an assistant invent references?
Because it is completing text rather than looking something up. Asked for a source, a model produces the most plausible-looking one, and a plausible-looking citation is exactly what a wrong one looks like.
This is not fixed by asking it to be accurate. It is fixed by giving it somewhere to search, so that the answer is assembled from a retrieved passage instead of from recall. The verification habits in Verifying AI financial data still apply afterwards, because a retrieved passage can still be summarised badly.
What does a documentation server return?
Titles, links and quoted fragments. Artha's exposes a single tool, `search_artha_docs`, over three sets of documents: the financial glossary, the explainer wiki, and published research.
The quoted fragment is the part that matters. An assistant that answers with a link alone is still paraphrasing from memory; one that answers with the passage it used lets you compare the claim against the words it came from. That comparison takes seconds and catches the errors that matter.
The single tool, the three document sets it searches, what it cannot reach, and the setup steps.
What can it not reach?
No portfolio, no watchlist, no alerts, no user record of any kind. It publishes no tool that would return them, so there is nothing to ask for.
That is a structural boundary rather than a policy one, and it is the reason Artha runs two servers instead of one endpoint with a public mode. Anything that needs your account sits behind consent on the other server, described in What is an MCP server. This one is public because it has nothing private to protect.
How do you add one?
Copy the endpoint and add it to your assistant as a custom connector. There is no sign-in step, no key and no cost, which also makes it the easiest way to see whether connector-based answers are worth the trouble before connecting anything that touches your account.
Then ask a question you can check — a definition, a convention, a published note — and open the link it returns. An assistant that cites a source you can open is a different tool from one that sounds sure.