LLM4UDT

A local AI powered web app that generates and optimizes Rockwell ControlLogix User-Defined Data Types (UDTs) for export to Studio 5000 as .L5X files.

Three input modes:

1. Natural language — describe a UDT in plain English, a local Ollama model extracts the field list, the app validates and emits an L5X.
2. L5X attachment — drop in an existing UDT or full-controller export, get back an optimized version with members reordered for minimum memory footprint and BOOLs bit-packed.
3. CSV / Excel batch — one tab or row-group per UDT, output is a zip of L5X files.

---

Why this exists

This is just a experimental extension of L5X UDT Optimizer. It leverages a local LLM to generate Rockwell Studio 5000 UDTs directly from NLP. This repo has achieved its intended purposes.

---

Install

Requires Python 3.10+ and (for mode #1) Ollama running locally with at least one model pulled.

pip install -r requirements.txt
ollama pull phi4          # or any model you prefer

Run

python3 app.py

Opens on http://127.0.0.1:5003. The app hard-binds to localhost and refuses to start on a non-loopback interface — there's no auth and it reads/writes engineering source files. If you really need a different host, set LLM4UDT_HOST=127.0.0.1 (or run it behind a reverse proxy with auth in front).

Optional environment variables:

Variable	Default	Purpose
LLM4UDT_HOST	127.0.0.1	Bind address. Only loopback values accepted.
FLASK_DEBUG	0	Set to 1 to enable Werkzeug debugger (RCE).

---

Usage

Mode 1 — Natural language

Open Settings (⚙️), pick an Ollama model, then in the chat box:

Create a UDT named MotorStatus with: run_status, BOOL, indicates if motor is running; fault_code, DINT; speed_setpoint, REAL; alarm_bits, BOOL[64]; phase_currents, REAL[3]

Field syntax is strict — fields separated by ;, values within a field by , (name, type, optional description). Supported types: BOOL, SINT, INT, DINT, REAL, STRING, TIMER, COUNTER plus arrays TYPE[N]. BOOL arrays must be multiples of 32 up to 1024.

Mode 2 — L5X attachment

Drag in a .L5X file (single UDT or full controller export). The app reports issues found (sort order, unpacked BOOLs, invalid BOOL array sizes) and offers an Optimize button. Output is a download.

For a full-program L5X, only the <DataTypes> section is rewritten — programs, tags, modules, AOIs, and tasks pass through untouched.

Mode 3 — CSV / Excel batch

CSV columns (case-insensitive): UDT_Name, Tag_Name, Data_Type, Description. Rows grouped by UDT_Name. Excel: one sheet per UDT, sheet name = UDT name. Output: a zip of L5X files.

---

Optimization rules

Members are sorted into groups (then natural-key by name within each group):

Group	Contents
0	BOOL scalars (packed into SINT bit-fields)
1–3	SINT, INT, DINT scalars
4	All arrays (sub-sorted BOOL → SINT → INT → DINT → REAL → LREAL → LINT → STRING → other)
5–7	REAL, LREAL, LINT scalars
8	STRING scalars (including Family="StringFamily" user types like STRING_40)
9	TIMER, COUNTER, CONTROL, MESSAGE, AOI instances
10	Nested UDT scalars
11	Unresolved / unknown types

Byte sizes follow Rockwell layout: natural alignment per type, BOOL[N] stored as ceil(N/32) × 4 bytes, consecutive BIT members packed into a single SINT block, final struct rounded to maximum member alignment.

Unresolved types are never silently dropped — they're carried through verbatim with a warning, so a UDT referencing a type defined elsewhere in the program won't lose members.

---

Tests

python3 test_core.py

29 regression tests, no external dependencies (Ollama is stubbed). Exits 0 on success.

Test areas:

• Engine-consistent sort in the program analyser (no stale local sort tables).
• BOOL[N] rounds up, never down — BOOL[33] becomes BOOL[64], not BOOL[32]. BOOL[>1024] is a blocking error, not silent truncation.
• LLM JSON fence parsing tolerant of CRLF, missing language tag, uppercase tags, missing trailing newlines.
• Stored-file cache is bounded (capped + TTL-evicted) and thread-safe.
• Localhost-only enforcement at startup.
• Upload directory purged at startup and per-request.
• String-family UDTs (STRING_40 etc.) sort with native STRING, not as arbitrary nested UDTs.

---

File layout

File	Purpose
app.py	Flask routes and request handling
model_UDTGen.py	Ollama → JSON tag extraction
Validator_UDT.py	Name sanitization, type validation, BOOL array sizing
L5XGen_UDT.py	L5X XML generation, member sort key, type sizes
L5XOpt_UDT.py	UDT optimization, alignment math, full-program rewrite
Attach_L5Xanalyzer.py	Detects L5X target type (DataType / Controller)
Attach_UDTVerify.py	Issue analysis for a single attached UDT
Attach_ProcessL5XProgram.py	Issue analysis + rewrite orchestration for full programs
Attach_ProcessCSV.py	CSV batch → zip of L5X
Attach_ProcessExcel.py	Excel batch → zip of L5X
Attach_UDTBatch.py	Shared batch zip builder
test_core.py	Regression test suite

---

Known limitations

• Single-process only. State (stored_files, known_udts, active_model) lives in module globals. Multi-worker deployment would need a session store.
• AOI size estimation only considers Parameters, not LocalTags. For AOIs with significant local storage the estimate will be low.
• LLM output is best-effort. If the model returns malformed JSON the request fails cleanly with a hint to rephrase — there's no retry loop.