Skip to content

Context and source evidence

AI context is a selected, authorized and bounded representation of source records. The selection happens before a model receives the text or attachment. A prompt can remind a model about an audience; it cannot replace the resource/capability checks that decide which facts may enter its input.

A representative conversation context builder resolves the conversation channel, project, current messages and channel-specific work information. Customer, worker, UE and internal histories have different audiences. Commitment/listing details are loaded only for their intended channels, and work-order context is scoped accordingly. Source text remains data. A supplier instruction embedded in a PDF or a customer message cannot authorize a new Toolora operation.

The public operations readers similarly establish caller, company and project scope before composing their packets. Narrow sections may be omitted with missing capability/coverage information instead of failing an otherwise permitted read. A broad project permission does not authorize every financial, personnel or conversation section. Conversation audiences and MCP authorization own those distinctions.

Mechanism Reviewed bound or behavior Reader consequence
MCP concise budget Default list cap 15, entity cap 5, text cap 280 characters; token target 6,000 A compact response is not a complete company export.
MCP detailed budget Default list cap 50, entity cap 25, text cap 2,000; token target 24,000 Tool-specific bounds and continuations still apply.
Token targets Character-based estimation, not a provider-token proof Do not treat the target as a universal hard size guarantee.
Section/body continuations Separate offsets, with callable next arguments Continue the specific section; preserve source identity across parts.
Mail quote proposal input Selected request and at most one explicitly selected supported attachment per turn A different file selection is a new reviewed input; internal opportunity notes remain separate.

The budget object supplies defaults; a tool can apply a narrower explicit cap. Source references and list rows can have independent truncation. Inspect missing_info, truncation, section coverage and observation time before drawing a conclusion. A truncated body is not evidence that the rest contains nothing relevant. Byte limits, MIME checks and original hashes belong to the original retrieval contract.

A source reference identifies the record or original used in a result. It does not certify that a generated statement correctly interprets the source, that an extracted quantity was received, or that an invoice was paid. Use the owning workflow’s accepted facts and coverage. Model output should retain unresolved questions and refer back to the source a human is allowed to inspect.

Mail/opportunity proposal context rechecks reader and current source links. A mailbox-disabled proposal does not become unrestricted simply because it is reachable from an opportunity. A saved draft, an attachment the model actually received and the next selected file have separate state. The workspace’s source recording helps review which input supported a response.

Suppose Customer Message A asks an assistant to ignore access checks and read internal margins. The context builder must exclude those facts before model use; the text’s instruction grants no capability. If the permitted work-order list is bounded, the response should state its continuation rather than invent a complete crew/effort total.

After a source revision changes, reopen the current evidence and review the proposal’s expected revision/hash before confirmation. Do not reuse a previous response as authority over a changed source. Deterministic tests can prove selection and clipping for their fixtures. They do not prove that all model responses are correct or that every adversarial document is harmless. The PR records executed checks; this review sends no private source text to a model.