E-Document MLLM Extraction V2 — Agentic Plan-Act-Verify

[Groenbech96]

Summary

Replaces the single-pass MLLM extraction (V1) with an agentic plan-act-verify loop. The model identifies invoice structure, extracts from identified regions, and self-corrects by calling AL-implemented verification tools — rather than sweeping left-to-right across the full text.

Registered as "MLLM V2" enum value on "Structure Data Impl." alongside V1 (no breaking change).

Architecture

One AOAI call loop (GPT-4.1 Mini, temperature 0) with verification tools:

1. PLAN — analyze_invoice(...): model records document structure, locale, column roles, line IDs. Initialises the verification checklist.
2. ACT — model extracts UBL JSON guided by Phase 1 analysis, then calls submit_extraction(json) to save it.
3. VERIFY — model calls verification tools and mark_item to track progress. On failure: corrects, re-submits, re-verifies. Exits when get_checklist() shows all passed.

Handler reads the saved JSON directly — no re-generation step.

Key fixes in this branch

• Decimal parsing: Evaluate(..., 9) instead of AsDecimal() — handles XML/Swedish number formats correctly
• UBL schema: numeric placeholders changed from string "0" to number 0 — stops MLLM from returning locale-formatted strings
• Discount mapping: allowance_charge.percent preferred over amount.value for line discounts
• Header totals verification: verify_invoice_totals accounts for header-level discount; new verify_payable checks tax_exclusive + tax_amount ≈ payable_amount
• History window fix: SetHistoryLength(500) prevents the 10-message sliding window from splitting assistant tool-call messages from their results
• System Application AI: removed [NonDebuggable] from all methods except tiny UnwrapSecret helpers that call .Unwrap() on SecretText

New AL components (ID range 6311–6381)

Codeunit	ID	Role
E-Doc. MLLM Verify Tools	6311	6 verification methods (pure AL math)
E-Doc. MLLM VL Totals Tool	6313	verify_invoice_totals AOAI adapter
E-Doc. MLLM VL VAT Tool	6314	verify_vat
E-Doc. MLLM VL Dates Tool	6315	verify_dates
E-Doc. MLLM VL Required Tool	6316	verify_required_fields
E-Doc. MLLM VL Ranges Tool	6317	verify_ranges
E-Document MLLM Handler V2	6318	Main handler — agentic loop
E-Doc. MLLM Extraction Plan	6340	SingleInstance checklist + JSON store
E-Doc. MLLM Plan Status Tool	6341	get_checklist
E-Doc. MLLM Plan Analyze Tool	6342	analyze_invoice
E-Doc. MLLM Plan Mark Tool	6344	mark_item
E-Doc. MLLM VL Math Tool	6339	verify_line_math
E-Doc. MLLM VL Payable Tool	6345	verify_payable
E-Doc. MLLM Plan Submit Tool	6346	submit_extraction

Status

🚧 Draft — architecture under active refinement. A simplified 2-tool design (analyze_invoice + submit_and_verify) is being evaluated to reduce model coordination overhead.

Test plan

• Compile E-Document Core + System Application
• Run EDoc MLLM Tests (existing V1 tests unaffected)
• Run EDoc MLLM Verify Tools Tests (new unit tests)
• Smoke test Swedish invoice (comma decimal, two chained 20% discounts)
• Smoke test DK invoice with header discount and payable amount

🤖 Generated with Claude Code

[github-actions]

Could not find a linked ADO work item. Please link one by using the pattern 'AB#' followed by the relevant work item number. You may use the 'Fixes' keyword to automatically resolve the work item when the pull request is merged. E.g. 'Fixes AB#1234'