Loading the knowledge base…
Inbox agent
Write a ticket the way a tenant or owner would. The agent either replies automatically, grounded in the knowledge base, or hands the ticket to a person and says why. Safety, legal and prompt-injection checks run on the raw text, so the model can never talk the agent out of an escalation.
Pick an example or write a ticket, then run the agent.
Knowledge search
The same question through three retrievers: BM25 (word matching), vector (character n-gram embeddings, so typos and compounds still match) and hybrid (both, merged by rank fusion). The agent tool also decides whether the best hit is good enough to answer from at all.
Browse the knowledge base ( documents)
Evals
Hand-labelled data, scored live in your browser. The numbers are regression baselines on small sets, not benchmarks, and the failures are shown rather than tuned away.
Escalation: does the right ticket reach a human?
37 tickets, each labelled "needs a human" or "routine". System: keyword rules plus the deterministic policy, no LLM.
| Ticket | Label | Agent |
|---|
Retrieval: is the right section ranked first?
27 questions the knowledge base answers, and 7 it cannot (those should be refused).
| Retriever | Hit@1 | Hit@3 | MRR@5 | Relevance AUC |
|---|
LLM router
Providers fail: rate limits, overloads, timeouts, bad keys. Choose how two simulated providers behave and watch the real router retry with backoff, fall back, and open a circuit breaker after repeated failures.
Settings for the demo: 3 attempts per provider, backoff from 300 ms with jitter, 1.5 s timeout, breaker opens after 2 exhausted requests and stays open for 10 s. Providers are simulated; the router is the real code.
What happened
- No requests yet.