Open source · built on free tiers · every number measured
AI customer support that only says what your docs say.
One script tag adds a support chat to your site. It answers from your own content with citations, hands off to a human when it should, and shows you exactly how accurate it is.
Free plan, no credit card. Source and architecture notes on GitHub.
Scripted preview
01
Add your documents
Paste your website URL or upload PDFs. Pages are crawled, cleaned, chunked and indexed in a vector store, scoped to your workspace.
02
Paste one line on your site
A single script tag renders the chat widget. Answers stream in and every claim carries a citation to the page or document it came from.
03
Humans take over when it matters
Low confidence or a request for a person pages your dashboard in real time. You reply in the same thread; the visitor never leaves.
The dashboard
Conversations, handoffs, cost and latency, and the questions your documents could not answer. Real data from the demo workspace, not a mock-up.

Built to be checked
Support bots fail quietly. Groundline is designed so that failures are visible and counted.
Grounded or silent
If the retrieved context doesn't support an answer, the assistant says so and hands off instead of guessing. Citations are validated after generation, not trusted from the prompt.
Measured in public
A fixed test set per workspace, scored on every run: accuracy, citation hit rate, correct refusals, over-refusal. Published at /evals when the owner opts in.
Multi-tenant by construction
Every query is scoped to a workspace at the data layer and inside the vector index. Isolation is tested, not assumed.
Humans stay in the loop
Low confidence pages a person. Tools that change something (tickets, bookings) wait for a teammate's approval. Costs and latency are visible per message.