YAU.POC/yau-plant-assistant
2
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
yau-plant-assistant/REQUESTS.md
Claude 41c4cbbbdc Correct H29 and the retention framing: the query window is always one week 
The query window is a rolling seven days by design, and stays that way against
the real SQL historian - it is not parsed out of the question. Two things I
wrote assumed otherwise.

H29 was wrong in two ways. It demanded the words "retention" and "seven days",
and it BANNED "no records found" - which contradicts CLAUDE.md, where that is
the required wording for zero rows. With a fixed one-week window June genuinely
has zero rows in what was queried, so the phrase is correct rather than evasive.
It also described the missing piece as question-window parsing, which is not a
gap but the design.

Rewritten to pin the risk that actually exists: SUBSTITUTION. A question naming
June must never be answered with this week's figure wearing June's label.
Answering "there were 14" would be exactly that, and nothing downstream could
catch it. The observed answer refuses the substitution and states the window it
used, so the case now passes on its merits rather than being red by default.
must_not is the only half run_eval enforces, so the banned phrases are ones
that appear only on a substitution.

outside_retention is reframed as what it is: a GUARD, always false while every
caller asks for seven days, there so a caller who later passes a longer window
cannot get an empty result that reads as "nothing happened".

REQUESTS.md and the removal guide now say that extending retention on its own
is inert - the window would still ask for a week of a longer history. Both
halves are needed, or neither.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-08-31 16:49:02 +10:00

13 KiB
Raw Blame History

Outstanding requests — WRPS Plant Assistant

One thing is still needed from other people, of the three originally raised. Nothing else is blocking: as at 28 August 2026 the assistant runs end to end on lin001 with a real model — the question box, name resolution, the data translator, document search, all four answer lanes, the rulebook, the working panel, the logbook and the document library are all live and were exercised by hand on the host.

Items 1 and 2 are both delivered and closed — they are kept below as a record of what was asked for and what arrived. What is still missing is real plant data. That is the request furthest outside this project's control, and it cannot be hurried at the end.

Item 3 needs the owner of the imh historian, who is not yet identified. Contact for items 1 and 2, both closed: Daniel Watson (daniel.watson@yokogawa.com).

# Request Needed for Blocks
1 Azure OpenAI account, 3 deployments The two AI steps — classifying the question, wording the answer Delivered 27 Aug 2026 — closed
2 Three public DNS records + DC pinpoint records Reaching the assistant by name from a control-room PC Delivered 27 Aug 2026 — closed
3 Read-only login to the imh historian Real plant figures instead of stand-ins Phases 4, 5; every data answer

1. Azure OpenAI account — DELIVERED

Closed 27 August 2026. The resource is live at yau-dem-oai.openai.azure.com, API version 2024-10-21, and lin001 reaches it outbound over HTTPS. NO_LLM_STUB=false on the host: both AI steps — classifying the question and wording the answer — run against the real model, and the documents were indexed with real embeddings the same day. Switching it on surfaced eight failures the stand-in had hidden; all eight are fixed and pinned by cases in eval/testset.jsonl, which grew from 67 to 75.

Two things are worth recording, and the second needs a decision.

  • The deployments as built are not the deployments as requested. The host runs CHAT_DEPLOYMENT=gpt-4o and CHEAP_DEPLOYMENT=gpt-4o — the same model on both lanes — with text-embedding-3-small for embeddings as asked. The request was for a GPT-5-mini-class model on the cheap lane specifically so that classification, entity extraction and tool selection would not bill at flagship rates.
  • The cost estimate below therefore no longer holds. ≈US$35/month at 3,000 questions was computed on three calls per question with two of them on a cheap model. With both chat lanes on gpt-4o, the same traffic bills materially higher. Either deploy a cheap-lane model and repoint CHEAP_DEPLOYMENT, or re-cost the budget ask before the demo period. The US$50 monthly budget alert should be confirmed as actually set, since it is now the thing that catches this rather than the estimate.

The original request follows, unchanged.

Action

Create an Azure OpenAI resource in the existing subscription and deploy three models on it:

Deployment Model
Small chat GPT-5-mini class or equivalent
Flagship chat GPT-5 class or equivalent
Embeddings text-embedding-3-small

Hand over, via a secure channel: the endpoint URL, an API key, and the API version. Confirm outbound HTTPS from lin001 to the endpoint is permitted.

Pay-as-you-go. No fine-tuning and no provisioned throughput — both bill a fixed hourly amount whether used or not, and neither is needed.

Cost

Consumption-based, billed on usage only. No licence, no commitment, no minimum term, cancellable at any time. Estimated on three model calls per question at published list rates:

Usage Questions / month Estimated
Occasional demos 500 ≈ US$6
Expected — build, testing and demos 3,000 ≈ US$35
Heavy daily use 10,000 ≈ US$120

One-off cost to index the document set: under US$1.

Requested: a US$50 per month budget, approximately US$250 for the demo period, with an Azure budget alert set at that figure.

Rates are published list prices and should be confirmed at approval. This is a demonstration figure; an operational plant deployment would be sized and costed separately.

What it is for

The assistant runs end to end today except for the two steps that need a model: classifying each question, and wording the answer. Without the account those two cannot be built, tested or reviewed, and the 67-question acceptance test cannot be run at all.

(As delivered: both steps now run. The acceptance test is runnable and has grown to 75 cases, but has not yet been run as a formal Phase 8 gate — that is a task inside this project, not a request of anybody.)

2. Three public DNS records, plus pinpoint records on the DC — DELIVERED

Closed 27 August 2026. All three A records resolve to 20.211.144.151, and the DC carries the pinpoint record ai.yokogawa.tech → 10.0.0.17. Nothing further is needed from Dan for this item. Two things are worth recording:

  • api.yokogawa.tech and cube.yokogawa.tech have no pinpoint record, so they do not resolve from inside the VNet at all. That would have broken the operator demo, because the browser called the API by that hostname. Rather than ask for a fourth record, Caddy now routes /ask under ai.yokogawa.tech to ai-api and the page is same-origin. No further DNS is needed for the demo. cube.yokogawa.tech is for engineers on the public side only; ai-api reaches Cube container-to-container and never by hostname.

  • auth.yokogawa.tech has no pinpoint record either, and that was the defect this item missed. Confirmed 28 August 2026: it does not resolve from inside the VNet. Every Authelia-gated service redirects there, so a LAN browser reached ai.yokogawa.tech, got a correct 302 to the portal, and then died on DNS — the name resolved, the service answered, and the operator still saw "this site can't be reached". It went unnoticed because the device agents write to Influx over the /api/v2/write MFA bypass and never touch the portal; no browser had ever hit Authelia from inside the VNet. This is no longer needed for the operator path — on 28 August the console was made unauthenticated from cicore1 only (README §6, BUILD-AI-CONTAINERS.md §14) — but any other gated service browsed from the LAN will hit exactly this wall, so it is worth asking Dan for regardless.

  • The names resolving is not the same as the services answering. The Caddyfile blocks and the Authelia rules for ai, api and cube were applied on the host the same day; all four AI hostnames now return 302 to the auth portal. What is still unproven is the part no check here can reach: an operator on cicore1 — since 28 August, with no sign-in at all — getting an answer end to end.

The original request follows, unchanged.

What is needed

Three A records under yokogawa.tech, each → 20.211.144.151, the same treatment lf.yokogawa.tech already has:

Name Points at For
ai.yokogawa.tech 20.211.144.151 The operator's screen
api.yokogawa.tech 20.211.144.151 The assistant behind it
cube.yokogawa.tech 20.211.144.151 The data translator

And — separately, on the domain controller — a pinpoint record for ai.yokogawa.tech → 10.0.0.17. Azure hairpin means LAN hosts cannot reach the VM's public IP from inside the VNet, so without this an operator sitting at cicore1 cannot resolve the assistant at all. influx.yokogawa.tech already has exactly this treatment.

What it is for

All three services are running right now and none of them can be reached by name. Today they are opened down an engineer's SSH tunnel, which is fine for development and useless for a demonstration. No control-room PC can open the assistant until these records exist.

Caddy already holds the site blocks, each with import authelia, so sign-in and TLS are handled the moment the names resolve; certificates are issued automatically on first request. Access is governed by the existing HTTPS_UserAccess group — no new AD group is needed for this item.

Why raise it now

This is the item most likely to be forgotten, because nothing fails until the day somebody tries. Every container is healthy, every check passes, and the gap only appears when a person outside this project opens a browser. It also depends on two different people — public DNS and the DC are not the same change.

3. A read-only login to the imh historian

What is needed

  1. The server. yau-sls-poc-imh:1433 is configured and waiting; the host is still being built.
  2. svc_agent_ro created with SELECT on the agreed tables only — no DDL, no write access, no xp_ procedures. Password over a secure channel.
  3. The database name, and the agreed table names and key columns.
  4. NSG rule: lin001 → imh on 1433 only.
  5. Timestamp semantics confirmed: UTC or local, and DST behaviour. This one is not a formality — see below.
  6. Permission to set an application name on the connection, so DBAs can see who is connecting.

What it is for

Every figure the assistant produces today is a stand-in. Cube — the data translator — is running and answering questions against generated tables (32 alarm activations, 72 pump-downs, about 685,000 analogue samples over seven days), and every answer built on one carries a warning label all the way to the operator's screen. The definitions Cube uses are the real ones and will not change when the historian arrives; only the source will.

The stand-in is now keyed the same way the historian is. It was rebuilt on 2026-08-31 against WRPS/05-scada/modbus — CI Server item names, the real historisation groups, the real sample rates and the real retention. Three defects that had been left open closed as a result, and the schema conversation below is correspondingly narrower than it was: the shape is now right, and what remains is names.

Why the schema conversation is still needed

  • What imh calls these tables and columns. One item-keyed history table is the right shape; fixture.item_history is our guess at its name. Everything else — alarms, pump-down operations — is derived from it, so this is the only naming question that actually blocks anything.
  • Whether imh exposes CI Server's ALARM_HISTORY group at all. It exists on the server and is empty: every WRPS item was imported with alarming off and limits at 0, which 05-scada/modbus/README.md records as engineering judgement nobody has made yet. Alarms are derived from the PLC alarm word instead, which needs no configuration that does not exist. If someone does configure item alarm limits, we would rather use them.
  • An engineer must confirm Cube's first real numbers by hand before anybody trusts one. The stand-in asserts its own counts at load and cross-checks them two ways, but that proves the pipeline, not the plant. This is a person's time, not just a login.

A separate request: extend the historian's retention

Every WRPS history group is LIFE_TIME "1 weeks". The assistant therefore cannot answer a question about anything older than seven days — not badly, but at all. "How did last month compare with this one", "when did we last spill", "how many trips this quarter" are all outside reach, and they are among the questions an operator is most likely to ask.

The system now reports this honestly: a window reaching past retention is answered as "the historian does not go back that far", never as "no records found", because those are different answers and only one is true. But honest is not the same as useful.

Asked for: raise LIFE_TIME on WRPS_ONE_SEC, WRPS_THIRTY_SEC and WRPS_EVENT to at least 90 days, or confirm that seven days is a deliberate constraint we should design around and stop asking. Storage is the trade: the 5-second group is five items, which is roughly 120,000 samples per item per week.

Note

Extending retention on its own changes nothing. The assistant's query window is a rolling seven days by design and is not parsed out of the question, so it would go on asking for a week of a longer history. Both halves are needed to answer "last month": the historian has to hold it, and the query window has to become variable. Raise them together or neither — and if seven days is confirmed as the intended constraint, the window work should be dropped too rather than left as a half-built capability.

Also worth a decision while someone is in there: the repository's his_group.qli and the live server's export_his_group.qli disagree about sample rates — the file says 1 s and 60 s, the server runs 5 s and 30 s. The live server was taken as authoritative. The repository does not currently describe the running system, and one of the two should change.

What it unblocks

Phase 4 and the remainder of Phase 5.

Related, but not a request

Phase 9 — operator document management will need a second AD group, AI_DocPublishers, with direct membership (Authelia resolves direct membership only; a user nested inside a group is silently denied). That is deliberately held until Phase 8 passes and is not being asked for now — noted here only so it is not a surprise later.