Grounded answers only

A New Era of Course Assistance

Upload your material and let students ask anything. The assistant answers from your syllabus alone — every claim traced to the page it came from.

Start free → See how it works
Databases 301 · assistant
Is ef_search the same thing as ef_construction?
No — ef_construction shapes the graph while it is being built and is fixed once the index exists. ef_search is a query-time knob you can change per session. We drew the distinction on the second slide of week 4.
◈ lecture-04.pdf · p. 9
Why did my HNSW index get slower after I raised ef_search?
◈ lecture-04.pdf · p. 12

Powered by Claude

Anthropic

Grounded. Isolated.
Live the same afternoon.

Three uploads and a model choice is the whole setup. There is no prompt to write and no pipeline to configure.

lecture-04.pdfindexed
seminar-transcript.mdindexed
problem-set-02.pdfchunking…

Drop the files in

PDF, DOCX, Markdown. Eight background queues chunk, embed and topic-map everything while you carry on with your day.

Pick a brain

Claude Sonnet 5 for depth, Haiku 4.5 for speed and cost, Opus 5 when the questions are hard. The grounding rules never change with the model.

No prompt engineering. The grounding contract is enforced in the agent, not in a prompt you have to maintain.

topic analytics · last 14 days
TopicSourceQuestions Struggle indexUnansweredTrend
HNSW parameter tuninglecture-04.pdf218 4▲ 62%
Isolation levelslecture-06.pdf174 2▲ 30%
Write-ahead loggingseminar-transcript.md143 0— 4%
Query planner costsproblem-set-02.pdf96 11▲ 18%
Normalisation formslecture-02.pdf61 1▼ 9%

Eleven students asked about query planner costs and got nothing back. That gap is a missing handout, and you find out on Tuesday rather than at the exam.

Built so it cannot drift

Six guarantees, enforced
in code rather than prose

Cite or stay silent

A factual claim must originate in a search result. No result, no claim — the student is told the course does not cover it.

Retry before it gives up

Thin results trigger a rephrase and another search. The loop runs until the agent has grounding or genuinely has none.

One tenant, one world

Reads, writes, queues and logs are scoped per course — on the request path and again inside every background job.

Topics map themselves

Every upload is mapped to the topics it covers, so questions land against a structure you never had to build.

It speaks as the course

First person, no plumbing. A blocklist in English and Ukrainian catches "based on the retrieved documents" before a student ever sees it.

Two languages, one voice

English and Ukrainian are first-class. The tone and the grounding rules are identical in both.

Every rule above is a test in the suite, not a line in a system prompt.

0%
of claims traced to a passage
0
answers from training data
0
background queues
0
languages, EN + UK
Voice

It speaks as the course,
not about it.

The retrieval machinery is deliberately invisible. There is a blocklist of the phrases that leak the plumbing.

  • First person. "In week 4 we settled on 40" — never "the documents state that".
  • Pedagogical, within limits. It can connect two of your documents; it cannot import a third idea from outside.
  • Honest gaps. Nothing retrieved means the student hears that the course does not cover it.
blocked phrasing
Based on the retrieved documents, the context provided suggests…
As an AI language model, I don't have access to…
We covered this in the week 4 seminar — the short version is that ef_search trades latency for recall, and past 64 you stop buying recall.
◈ seminar-transcript.md

What changes on Monday morning

“The repeat questions stopped landing in my inbox. What arrives now is the genuinely hard ones — and the analytics tell me which week produced them.”

Maryna Kovalenko
Lecturer, Data Systems

“I had to be certain it would never invent an answer about our compliance module. The citation on every reply is what got it through review.”

Tomas Bergmann
Head of L&D, fintech

“Uploaded eleven PDFs on a Tuesday. Students were using it on the Wednesday. There was no third step.”

Olena Lysenko
Founder, bootcamp

The three questions everyone asks

Will it make things up?
It is built not to. A factual claim has to originate in a search result; the only permitted extension is connecting documents that were both retrieved. If retrieval comes back thin the agent rephrases and tries again, and if that still fails it tells the student the course does not cover it.
Can another course see my material?
No. Scoping is enforced twice — once on the request path and again on every background job, which carries its tenant explicitly rather than inheriting it. Reads, writes, queues and logs are all partitioned. See the security model →
Which model does it use?
Claude Haiku 4.5, Sonnet 5 or Opus 5 — your choice, per course. Change it at any time; the grounding rules do not change with the model.

Take control of your course

Upload your material, pick a model, share the link. Most courses are answering questions the same afternoon.

Start free → See pricing