How a turn works ​

A question passes six stages. Four of them can end it, and three of those four end it before any model is called.

1. The credential check ​

Synchronous, network-free, first. A question carrying a credential shape settles in the browser. See Credentials in questions.

2. Is it a question at all? ​

"Hello" carries no documented subject, so the gate below would score it at zero and refuse it — a correct verdict that tells the reader, on their very first message, that the assistant is broken.

So a greeting, a thank-you, a farewell and "who are you" are recognised here and answered from a template, with no model call. Only an input that is social and nothing else is claimed; "hi, how do I authenticate?" carries a subject and takes the normal path. See Social openers.

It runs after the credential check, so a greeting with a key pasted into it is still handled as a key.

3. Retrieval and the gate ​

The question is embedded once, scored against every chunk in scope, and fused with a keyword pass. The gate combines two channels into one score and compares it to a calibrated threshold. Below it, the turn ends: no model call, no generated text, and the panel shows the closest pages instead of an answer.

Retrieval is hybrid, and both halves always run. A BM25 pass over the chunk text — MiniSearch, over the text, the title and the breadcrumb, with the title weighted — and a cosine pass over the 1-bit vectors produce two ranked lists, and the two are merged by Reciprocal Rank Fusion: each list contributes 1 / (k + rank) to every id it names, and the sums are sorted. Ranks, not scores, which is what lets two channels with incomparable units vote in the same election. The fused pool is then re-ranked by dense cosine before the gate sees it.

With no embedder configured, or with one that cannot be reached, the dense list is simply empty: BM25 runs alone and the gate switches to its lexical threshold. That is a mode, not a failure — but it is an expensive one, and what it costs in recall is measured rather than guessed.

See The refusal gate.

4. The primed turn ​

Above the threshold, the host does something most retrieval loops do not: it puts the gate's own excerpts into the first message. The model starts with evidence rather than spending a step asking for it. Most turns therefore need one call, not two.

When that call does not come back, the answer ladder is what happens next: every other service the environment selected is asked in turn, and a turn where none of them answers settles as the passages retrieval already found rather than as an apology for them.

5. Tools, if the model wants more ​

search_docs, fetch_section, expand_section, list_pages, and answer. Two properties matter more than the list:

None of them accepts a path, a page set, or a scope. There is no argument in which a wider scope could be expressed, so scope is not a rule the model is asked to follow.

expand_section is the one that exists for a shape of failure rather than for a capability: an answer that begins at the end of one section and finishes at the start of the next. It takes an id the model already has and returns the section immediately before or after it, on the same page. Two per turn — enough to cross a boundary in either direction, not enough to read a page a chunk at a time — and it always costs a step, because a free one is a walk with nothing to stop it.

A site can add its own. A function tool from tools joins this list on every search step. DocPilot checks its arguments against the site's schema, runs it, and returns the reply as an observation with an id the answer may cite, such as tool:plan_limits:1. A site tool takes the arguments the site declared, and none of them is a scope either: the reply is the site's data, not a chunk. An action is never in this list. The answer offers it, and the reader runs it.

Observations are appended, never rewritten. Step two's prompt is step one's prompt plus new results at the end. This is what keeps a provider-side prefix cache valid across the steps of a turn; rewriting the head to save tokens trades a small saving here for a much larger loss there.

An excerpt is spelled out once per turn. A chunk two searches both return carries its id and title the second time, without the body the model has already read.

6. Validation ​

The answer is checked before it renders:

• Citations are compared against the set of ids the host emitted this turn. An id the model invented is stripped, and its marker with it; surviving markers renumber.
• Links are checked against the page set in the token stream, before rendering. An invented route is de-linked and left as plain text.
• A citation of a site tool's reply passes the same check: the host handed out its id this turn. Its source row has no link.
• An answer with no surviving citations is not shown. It becomes a refusal — an uncited answer is exactly what the guardrail exists to withhold.
• actions are checked against the site's own schemas. An action the site does not have, an argument that fails its schema and a duplicate are dropped, and at most three remain. None of them runs: the reader decides.

A low-confidence answer that is cited is kept and marked tentative. Confidence is the weakest signal in the system: a number the model writes about its own work. Deleting a checkable answer on that basis costs the whole turn and replaces something useful with a claim about the corpus that is not true.

Conversation memory ​

Three question–answer pairs travel, the questions verbatim. The last answer travels whole, up to 2,000 characters, because the next question most often points into it — "step 4", "make that use React". Each older answer travels as its summary: one or two sentences the model writes into a summary field of every answer, which the reader never sees. An answer with no summary — a thread reopened from an older archive, a model that skipped the field — falls back to its skeleton: the opening sentence, then the leading line of every block after it, never a cut inside a code fence, within 360 characters. Earlier questions collapse into one line naming their subjects, and only turns that actually completed are carried — an empty assistant message is not a neutral placeholder, and two refused turns ahead of an answerable one were measured turning it into a refusal.

The summary line is built from prior questions only, never from answer text. A memory slot the model authors itself, outliving the window, is a multi-turn injection channel the gate cannot see.

The gate keeps the same rule one step earlier. Its composed channel composes the follow-up against the last question that was answered, not the last one that was asked — and, when that turn was itself an ellipsis the composed channel rescued, against the question it followed as well. A refused turn is a question this corpus retrieved nothing for, and putting it in front of the follow-up anchors the one channel that could have recovered the turn to a known dead end. That is the same turn the window above drops, measured for the same reason.

Condensing is why a reader can quote: a summary or a skeleton names what an older answer was about and carries none of its wording, so a passage from the middle of a long answer is still not in context verbatim. Select a passage in an answer and it is attached to the next question as its own field — labelled as quoted text rather than as an instruction, and carried with that question one turn later, clamped shorter. Without it, a question about the fourth paragraph of a long answer arrives with that paragraph nowhere in context.