Skip to content
Fewer tokens, by design

Your AI gets the passage, not the pile.

Pasting your whole vault into a prompt burns tokens and trust. Altretta hands your model only the cited passages that matter, over MCP, and the saving is measured rather than promised — 94.3% fewer input tokens than pasting everything, on 24 pre-registered questions you can read below.

The headline

Two numbers, because the answer depends on the baseline.

A percentage saved means nothing without the thing it is saved against. The two defensible baselines disagree, so both are stated — and the conservative one is the smaller.

94.3%vs pasting your whole vaultMedian over 24 questions. Range 92.3%–95.2%.
26.5%vs an assistant that already picks the right filesMedian over the 18 answerable questions. The sceptic's number — and it goes negative on single-note lookups.
74.7%over an 8-query conversationBilled the way a chat API bills: the baseline pastes the vault once and carries it; Altretta's passages accumulate turn by turn.
Every question

The 24 questions, one row each.

Pre-registered before the measurement and never derived from our own engine's output. The oracle column is blank for questions the vault cannot answer — establishing an absence means reading everything, so the oracle degenerates to the whole vault.

QuestionShapeWithout Altretta (whole vault)Oracle filesWith AltrettaSaving vs whole vaultRetrieval
Who is Calcifer and what kind of being is he?Single note23,0163611,20194.8%grounded · pass
Which character does the Ashitaka Sekki theme portray, and what curse is he carrying?Cross-note23,0238281,42793.8%grounded · pass
List every character who has their own note in this vault, and the film each one belongs to.Vault-wide23,0245,0621,10295.2%grounded · n/a
Which voice actor played No-Face in the English dub?Absent23,0161,10095.2%grounded · pass
Which award did Spirited Away win in 2003, and what box-office record did it hold in Japan?Single note23,0277821,77292.3%grounded · pass
Who directed Grave of the Fireflies and who composed its score?Cross-note23,0171,9751,23694.6%grounded · pass
Give the runtime in minutes of every film in this vault.Vault-wide23,0167,8201,52593.4%grounded · n/a
What was the production budget of Princess Mononoke in yen?Absent23,0171,15495.0%grounded · pass
What does the title Ashitaka Sekki mean in English?Single note23,0174121,47893.6%grounded · pass
Which Ghibli films in this vault were not scored by Joe Hisaishi, and who scored them instead?Cross-note23,0271,5991,17994.9%grounded · pass
Name every named piece of music mentioned anywhere in this vault and the film it belongs to.Vault-wide23,0226,3111,22594.7%grounded · n/a
In which musical key is One Summer's Day written?Absent23,0151,59293.1%grounded · pass
Where and when was Joe Hisaishi born?Single note23,0146711,66392.8%grounded · pass
Who founded Studio Ghibli, and what was each founder's role?Cross-note23,0193,0391,40793.9%grounded · pass
Which people are named somewhere in this vault but do not have a note of their own?Vault-wide23,0225,9501,25494.6%weak · n/a
What is Joe Hisaishi's real birth name?Absent23,0151,65392.8%grounded · pass
In what year did the Ghibli Museum open and where is it located?Single note23,0205141,49393.5%grounded · pass
Which exhibits at the Ghibli Museum come from which films?Cross-note23,0172,0601,50993.4%grounded · pass
Which notes in this vault link to Studio Ghibli?Vault-wide23,0168,6001,24294.6%grounded · n/a
How much did it cost to build the Ghibli Museum?Absent23,0171,55293.3%grounded · pass
Which film inverts the food theme into starvation, and with what image?Single note23,0194841,17594.9%grounded · pass
Which two heroines of Princess Mononoke are described as opposing faces of female power, and what does each of them want?Cross-note23,0301,3281,46193.7%grounded · pass
For each of the six themes in this vault, list every film cited under it.Vault-wide23,0212,8851,10695.2%grounded · n/a
Which theme does this vault say appears in the largest number of Ghibli films, by count?Absent23,0241,20794.8%grounded · pass

Token counts are in cl100k_base, the public BPE every headline figure is reported in. Negative savings are kept, not hidden.

The method

Three arms, one honest comparison.

The claim is about input tokens per query — how much text has to be put in front of a model to answer a question about your notes.

ARM A1 — the whole vault (weak baseline)

Every note in the vault, once, each under a path header, plus the question. This is what a filesystem assistant with no retrieval does. The counterfactual the marketing copy names: "instead of your whole vault".

ARM A2 — oracle file selection (strong baseline)

An assistant with no passage retrieval but perfect knowledge of which files hold the answer, reading each whole. Stronger than any real search: it never picks the wrong file and is charged nothing for searching. The conservative figure a sceptic demands.

ARM B — with Altretta

The exact bytes the MCP tool aingle_ground returns at its default k=6, plus the question. Not an approximation — the full payload, including every citation and provenance anchor. The overhead of provenance is charged to us, not hidden.

What is excluded from every arm

The system prompt, the MCP tool definitions, and the model's own replies. Tool definitions are a real, recurring context cost that only the Altretta arm pays, and this benchmark does not charge it — a known bias in our favour, stated here rather than left to be discovered.

Reproduce it yourself

The protocol runs over the Studio Ghibli demo vault, which ships inside the app — no private data, nothing to take on trust. Install the Altretta skill and run the same questions.

Get the Altretta skill

Spend tokens on the answer, not the context.

Altretta is free to start, runs on your machine, and connects the AI you already pay for.