Five separate repos that actually call each other: an agent that retrieves, a gateway governing its model calls, an eval harness grading its answers, a dashboard showing the grades, and a database that remembers every run. This page installs each project's own published wheel, the exact artifact its CI built, and runs them together in your browser. The last one isn't in your browser at all: it's eval-history's hashed static archive of a real FastAPI + Postgres service, and this page really reads it. The arrows below are function calls, not a diagram.
Booting…
rag-eval-labllm-gatewayagent-grapheval-dashboard (at the end)eval-history (archive)
Four questions go through all five projects, ending with an eval_run.json you can open in the dashboard, and a verdict from eval-history's archive. Watch question 4: it isn't in the corpus, and that's the interesting one.
or ask it your own
What's a corpus? Just the pile of documents the system is allowed to look things up in: its entire world knowledge. That's the whole idea behind RAG: instead of letting a model answer from memory, you force it to answer from documents you chose. Which is exactly why the answer can then be checked: a model answering from memory can't be audited; one answering from your documents can.
So the stack handles any topic; you just have to hand it the documents. Ask about something that isn't in there and it won't know, which is the interesting case.
← swap the whole world it knows about
…or write your own, one id: text per line
Paste your own documents and the retriever rebuilds around them; the planets are just what the demo shipped with. Then ask a question about your corpus above.
What's real and what isn't, plainly. The retrieval, the grading, the gateway's auth/cache/cost accounting and the agent's graph are all genuinely executing, installed from each project's published wheel. The provider is llm-gateway's deterministic mock, which echoes its prompt rather than writing prose; that's what lets this run with no API key and no spend. So the grade is computed against the agent's own composed answer, not the mock's echo. Point the gateway at a real key and nothing about the wiring changes.