Foreslaa regel: Evaluer BCApps AI Test Toolkit foer custom eval-harness til LLM-features

custom/architecture/ai-test-toolkit-for-llm-feature-eval.md

@@ -0,0 +1,54 @@
+---
+bc-version: [24, 25, 26, 27, 28]
+domain: architecture
+keywords: [ai, llm, copilot, eval, ai-test-toolkit, evaluation, testing, accuracy, token-consumption]
+technologies: [al]
+countries: [w1]
+application-area: [all]
+---
+
+## Description
+
+Before building a custom AL test harness to evaluate an LLM-based feature's
+output quality — Copilot-style document parsing, chat, extraction, or any
+capability that calls an LLM and needs a reproducible accuracy measure —
+check whether Microsoft's BCApps **AI Test Toolkit** (`src/Tools/AI Test
+Toolkit/` in microsoft/BCApps, starting at codeunit 149044 "AIT Test
+Context") already covers the need.
+
+## Why
+
+The toolkit is a thin, model-agnostic wrapper: `GetInput()`,
+`GetExpectedData()`, `SetTestOutput()`, `SetAccuracy()`,
+`SetTokenConsumption()`. It does not call any LLM itself — your own test
+codeunit calls whatever model you use (Azure OpenAI, Anthropic Claude,
+anything) and reports the result back through the harness. Its only
+dependencies are System Application and Test Runner — no Azure OpenAI
+binding, no Copilot-specific requirement.
+
+Adopting it gives you, for free:
+- A standard `.jsonl` dataset format (`input` + `expected_output` per line)
+- Built-in accuracy and token-consumption metrics per test case
+- A results page inside Business Central itself, instead of a bespoke one
+
+What it does NOT give you: the actual evaluation logic. You still write
+100% of the comparison/assertion code that calls your feature and checks
+its output against the expected result. The toolkit saves scaffolding, not
+engineering effort — adopting it is close to free, so the bar for skipping
+it should be high.
+
+## Rule
+
+Before building a custom AL test harness for evaluating an LLM-based
+feature's output quality, check whether AI Test Toolkit already provides
+it. Skip it only with a documented reason (e.g. a genuine feature gap in
+the toolkit for your use case) — not merely because it is unfamiliar.
+
+## Verified 2026-07-22
+
+Source-level inspection of AI Test Toolkit v28.0.46665.48704 (extracted
+from a local BC sandbox artifacts cache) confirmed the claims above. Two
+commonly-assumed capabilities were checked and NOT found in the toolkit's
+actual AL source: an Azure-AI-Foundry-based "Red Team Scan," and any
+Azure-OpenAI-specific binding. Do not assume either exists without
+re-verifying against the version you target — BCApps evolves.