Architecture

The services on the edge PC, how data flows through them, and what stays vendor-neutral.

BooleX runs on one edge PC that the site provides, in its own network zone beside the control system. It is a set of containers started with Docker Compose (deploy/compose.yaml).

Services

Only the assistant publishes a port (8080/TCP, HTTPS outside development). Every service restarts on exit.

How an answer is made

1. Data arrives: live values and alarms through ingest (or from the site historian, or as imported files); code exports, documents, work orders and parts through the importers.
2. Code from every vendor is translated into one common code model: programs, rungs and networks, tags, alarms, interlocks and their cause chains. The assistant reasons only over this model and the historian interface, never over a vendor format or a specific database.
3. A question, or a stop, starts a diagnostic session. The assistant runs read-only tools (cause chain, sequence of events, trend, documents, work orders, parts, model replay) and builds the evidence card without a language model.
4. The language model, behind the provider interface, writes the answer from the tool results. Every statement is checked against its cited sources before it is shown; a statement without a source is dropped.

Language models

The model is configured in deploy/config/llm.json and reached only through the provider interface. In the production and offline-import profiles it must run on the edge PC or the plant network; a cloud model is refused at start-up. See Deployment profiles.

Focused agents (experimental, off by default). With "agents": true in llm.json, a Lead investigator hands each question's sub-questions to specialists by expertise: Process, Control logic, Drawings and documents, Maintenance and, where validated models are on, Behavioural models. One model plays every role; each specialist is offered only its own tools, and every claim it passes back must cite a source the tools returned. Rules choose the first specialists; when none match, the Lead plans with one short model call, and an unclear plan falls back to the normal answer with all tools. Up to challenge_rounds rounds (2 by default) let the specialists dispute each other's claims. Each question is limited to 6 investigation tasks, with the challenge rounds on top, and to agents_time_budget_s seconds of specialist work (20 by default, at most 120). The answer format and every check stay the same, and the audit trail records every task and finding. candidates_shown (3 by default) sets how many causes are shown in full and two_pass turns on the two-pass answer for any model. Leave the agents off until the benchmark shows they help your model.

Where things are kept