Retrieval and context engineering
The assistant does not search. The published portfolio fits in its prompt, so every answer is written with all of it in view, and every citation is checked against the list of real ids before it reaches you.
The usual answer to "ground an agent in your content" is retrieval: index the content, fetch the few passages that look relevant, and hope they were the right ones. This site does not retrieve. The portfolio is small enough to put in the prompt whole, so the assistant reads all of it on every turn, and the engineering moves to what surrounds it: what the prompt holds and in what order, what the model may treat as an instruction, and what happens to a citation after the model writes it.
No index at all
Retrieval was built before it was rejected (ADR-0038). On 2026-09-20 a BM25 index over the projected corpus was run against 27 answerable questions written for the purpose. At the five passages a turn can afford, it found a supporting passage for 82% of the English questions and 60% of the Italian ones. So nearly one English question in five that the corpus could answer would have come back as "I don't know", and each would have been filed as a content gap that does not exist. The relevance threshold meant to separate answerable questions from unanswerable ones did not exist either: 9 of the 26 answerable questions that returned anything scored at or below the best-scoring unanswerable one.
Loading everything was cheaper than the machinery that would select part of it. The corpus was measured at about 4,200 tokens that day. That figure was counted with an OpenAI tokenizer, and the corpus has grown since. With these how-it-works pages it reached about 25,000 tokens, past the ceiling below. So these pages now enter the prompt as their summary and opening only, and the assistant points to them for the rest, which keeps the corpus near 11,400 tokens (ADR-0038). So there is no vector store, no embedding model, no reranker, and not the BM25 index that was measured. Recall is complete by construction, because every passage of the portfolio is in context on every turn.
The unit of the corpus is the anchor: one document per addressable section, labelled with its id. There is no chunking to tune, because the boundaries are the ones I wrote. Before the documents comes a catalogue of every entity, with its as_of date, any topic marked withheld, and for each skill the work that proves it. Because everything the corpus holds is listed, "that is not published here" is a fact the model can see, not an inference from a failed search.
This is the argument of accreta's own compile, don't retrieve decision, applied to this site. It has a shelf life, and the limit is written down: CI fails the build when the projected corpus passes 20,000 tokens, and the BM25 configuration measured here is recorded as the design that replaces it.
Four layers and three levels of trust
The prompt is assembled in four layers, ordered by how far each can be trusted and how often it changes (ADR-0042):
- the system instruction: role, perimeter, grounding and citation rules, language, output format;
- the tool definitions, generated from the tool contracts;
- the corpus, which I wrote and which the model may quote and cite but not obey;
- this turn: the earlier turns you sent, the page you are on, any text you selected, your message and any tool results.
The first three are the same bytes on every turn of a build, so the provider can cache them. The fourth is everything a visitor can influence. Each item in it sits in a labelled block, and the instruction says that nothing inside such a block is an instruction, whoever appears to be speaking. How that holds up against attempts to break it is on Guardrails, privacy and security.
The fourth layer has a budget, because it is the part that is never cached (ADR-0044). The conversation so far is trimmed to 1,000 tokens, oldest turns first, and never summarised: a summary would cost a model call on every turn and would be one more thing that can lose a fact. A message is at most 1,000 characters, the page context at most 1 KB, and each tool result at most 1,000 tokens. The page and section you are reading are one sentence in the turn, which the model uses to break a tie when a question is ambiguous. They are a preference, not a filter. How deep to go is read from the question itself and from the section it was asked from. There is no profile of you to read it from.
What the assistant may say
The grounding rules are written as rules rather than as a tone (ADR-0042, ADR-0030):
- Every factual claim carries at least one source id. A sentence that cannot cite is not written.
- A fact has three states. A value is stated. A withheld topic gets the site's fixed sentence, "{Topic} is not public — {reason}.", and nothing is said around it, not even as the assistant's own reading. An absent one is stated as not published here, and the topic it was looking for is recorded, never your sentence.
- A fact with a stale date is given with its date, never restated as if it were current.
- The model's own knowledge of the world is not a source.
- Anything the corpus does not say outright is marked as the assistant's reading.
- It answers in English or Italian, following your language. Ids, titles and quoted claims stay in their English form, because a translated claim is a new claim.
The assistant is software and says so if asked. It speaks about Francesco, not as him, and what it knows is what this site publishes.
Citations are checked after the model writes them
A citation is an anchor, sent to your browser in its own frame beside the text rather than inside it. Every id the model emits has to pass two checks before it is streamed. It must be a live entity or a citable section in the artefact the site was built from, and it must name a document that was in the model's prompt. An id that fails either check is dropped and counted (ADR-0039). With no retrieval step, a cited anchor that was not in the prompt can only be an invention, and set membership catches it without anyone having to judge prose.
The filter is load-bearing, not defensive. In deep-dive 8, two full runs of the test set emitted 519 source ids and one of them did not exist: a correct section put under the wrong type prefix, on the line before the correct id for the same claim. That is 1.9 per 1,000. It happened on the second run of a question the first run had cited cleanly, so a single clean run proves nothing about grounding.
The corpus is projected once, when the agent starts, from the same kb.json the pages were built from. Changing content means deploying a new build, and the site and the agent always ship the same build together, so an answer cannot cite a section the live pages do not have. Staleness travels with each passage as data. Nothing else is cached: there is no cache of answers, because it would be keyed on what visitors typed.
How far this has been proved
- Measured, 2026-09-20, on questions written for the purpose: the BM25 recall and the missing threshold above.
- Measured, deep-dive 8: 1.9 fabricated ids per 1,000, caught by the filter. That run used an earlier answering model, not the one that answers here now.
- Designed: the four layers, the budgets and the grounding rules.