Split page by page, never across a page boundary, so a citation can name one.
→ page 7
Retrieval you can check
Ask your documents. See where every answer came from.
Add your PDFs, lectures, transcripts and links. Every sentence of every answer carries a marker that opens the exact page, paragraph or second it came from. When the answer is not in your sources, you get told that instead of a guess.
Your notebooks are private to you
What has to happen before a write is committed?
A minority cannot commit on its own, which is what stops two halves of a split cluster from both accepting writes. A write is committed only once a majority of nodes have acknowledged it[1]
...the leader appends the entry to its own log and waits. A majority of nodes have acknowledged it before the entry is considered committed and applied to the state machine...
The part most tools skip
It would rather say no.
Before a single word is written, the passages that came back are scored out of ten on whether they can actually answer the question, and asked what is missing.
Below the line, those gaps become new search terms and the search runs again, up to three rounds. The best round is what reaches the model, not the last one.
If no round clears the bar, you get the sentence beside this one. The strongest guard against a confident wrong answer is never asking for one.
I could not find this in your sources.
With the option to widen the search across every source, rather than a dead end.
Seven steps
What happens when you ask.
One vector lookup on your raw question is not good enough, and it fails in two places: the question is often a poor search key, and whatever comes back is normally accepted without anyone asking whether it is enough. Each step is switchable, so every stage has to earn its latency.
- 01
Translate
Your question becomes several in one call: a standalone rewrite with references resolved, a broader step-back question, its sub-questions when it is compound, and a hypothetical answer to search as prose. Your own wording is always kept alongside them, so a bad rewrite can never discard it.
- 02
Route
Vector search for what the sources say. Keyword search when the question turns on a quoted string, an error code or a name that embeddings blur. A read-only database query for questions about the collection rather than its contents.
- 03
Search
Every variant against every routed channel, concurrently. A channel that fails returns nothing rather than failing the question, and the trace records what was lost.
- 04
Fuse
Reciprocal rank fusion across the ranked lists. A passage that surfaces across many rephrasings is more likely to be the right one than one that surfaced once, and that agreement is the reason the variants exist.
- 05
Rerank
A cross-encoder scores the survivors against your original question rather than any rewrite of it, and keeps the top eight.
- 06
Grade
A small model scores the retrieved set out of ten for whether it can answer at all, and says what is missing. Those gaps become the next search.
- 07
Answer
Numbered context blocks in, markers out, mapped back to real chunks server-side. A marker pointing at a block that was never supplied is stripped rather than shown.
Five kinds of source
Five kinds of source, five kinds of citation.
A citation is only useful if it points somewhere exact, so each type is split along its own natural seam and carries the locator that suits it.
YouTube
Captions pulled per cue and merged into passages of roughly a minute of speech, with the video playable beside them.
→ 04:12 – 05:03
VTT and SRT
Your own transcripts, with speaker labels kept where the file has them.
→ 04:12 – 05:03
Web page
Fetched and reduced to the article text, with the heading path kept per block.
→ section, characters
Text
Pasted or uploaded, with character offsets preserved exactly.
→ characters 1200–2100
Web pages are captured when you add them, so the viewer still shows what the model read even if the site later goes down or blocks embedding.
Roadmap
A roadmap through the material
Ordered modules built from what your videos and transcripts actually teach. Each one is pinned to the timestamps where the concept is covered, and a concept with no pin does not make the list, so the roadmap cannot invent topics your sources never mention. Tell it whether you are new to the subject and it changes the granularity.
Podcast
A podcast from your notebook
Two hosts, three to ten minutes, written only from summaries of the sources you picked and told to introduce nothing else. The script sits beside the player with each turn linked to the sources behind it, so you can check what was said rather than take it on trust.
Your sources. Your questions. Nothing invented in between.
Add a PDF or paste a link, ask one question, and click the marker on the answer. That is the whole product in about a minute.