Multi-Plant Master-Data-Aware Purchase Order Extraction v2.0

Charles Dana · Monce AI · May 2026

saft.aws.monce.ai · Two plants live: Valdosta GA · Bordeaux FR

Abstract

Saft (TotalEnergies) operates two plants on this pipeline: Valdosta GA (9,425 SKUs, 22 customers — Verizon, Satair, Airbus, Boeing, Lockheed, Textron, Learjet, Safran) and Bordeaux FR (43,332 SKUs, 2,990 customers — Saft Ferak a.s. is the demoed customer on PO P2502253, 264 lines, €448,650.50). Each PO references customer-side or manufacturer part IDs that must be reconciled against the plant's ERP master data. We formalize the matching step as a plant-aware cascade over a weighted SAT classifier (Snake v5.4.6) and four deterministic fallbacks, prove that the output is order-independent, auditable, and invariant to noise in the VLM-extracted description, and report measured accuracy on the full live PO sweep: 274 / 275 lines tier-1 exact (99.6%), $0.69 aggregate Bedrock cost, 412s wall.

1. Problem statement

Let a Purchase Order be a set of n line items

where each ℓ_i = (m_i, c_i, d_i, q_i, u_i) with

• m_i ∈ Σ* — the manufacturer part ID as extracted by the VLM,
• c_i ∈ Σ* ∪ {ε} — the customer part #, optional,
• d_i ∈ Σ* — the free-text description,
• q_i ∈ ℤ_≥0 — the quantity (EA, LB, FT, P, etc.),
• u_i ∈ ℝ_≥0 ∪ {⊥} — the unit price (USD or EUR), often ⊥ on framework POs.

Each PO carries a plant routing p ∈ P = { valdosta, bordeaux }, determined at Stage 0 from the supplier address. Stage 4 returns an assignment σ : L → M_p ∪ {⊥} where M_p is the plant-specific master data:

• |M_valdosta| = 9,425 SKUs (OM-Material extract)
• |M_bordeaux| = 43,332 SKUs (Master Data dump from Stéphane, 2026-05-29)

2. The Snake basis (per plant)

Snake (v5.4.6, cf. Dana 2024) builds a weighted SAT classifier from (text, label) pairs. Three models are trained per plant — six models total in production:

• article_matcher: free-text → Saft SKU. Trained on (description, number) pairs from M_p plus synonym rows for known aliases (mfr-PN ↔ SKU, customer-PN ↔ SKU).
• client_matcher: free-text → customer_id. Trained on (name + address + PO-number-prefix, customer_id).
• entity_matcher: confirms the receiving plant. Trained on (address, email, phone, VAT) pairs pointing to SAFT AMERICA / SAFT BORDEAUX.

Each model emits prediction with confidence γ ∈ [0, 1] and a human-readable audit. Training knobs (per internal playbook): n_layers=5, bucket=50, noise=0.25, oppose_profile=industrial, workers=5 for the 43k-class Bordeaux article model. The 9.4k-class Valdosta model uses n_layers=15 for max accuracy.

3. The plant-aware cascade

Given a line ℓ in plant p, define Σ_p : L → M_p ∪ {⊥} as the first-match cascade:

Σp(ℓ) = { k ∈ Mp : key(k) = m }                  (tier 1: exact on mfr PN)
        { k : key(k) = c }                          (tier 1: exact on customer PN)
        { k : key(k) ∈ synonyms(k) ∩ (m, c) }     (tier 2: synonym alias)
        article_matcherp.predict(m + " " + d)    if γ ≥ θauto  (tier 3: Snake)
        match_with_star(m, Mp)                    (tier 4: star-notation)
        fuzzy(m + " " + d, Mp, ≥ θfuzz)         (tier 5: Levenshtein)
        ⊥                                          otherwise

with θ_auto = 0.85 (the Snake confidence floor; lowering it has bitten us in past silent SKU hallucinations) and θ_fuzz = 0.80. The cascade terminates in exactly one of seven branches per line.

3a. Plant-specific opt-ins

The cascade carries one plant-conditional rule. Bordeaux accepts the line's customer_part_number as a candidate at tier 1, because Saft Ferak embeds the Saft SKU verbatim under the customer PN. Concretely, all 264 Saft Ferak lines have customer_part_number = "60-I" + saft_sku:

60-I105006   →  105006        (rondelle M6 A2)
60-I791111PCBA → 791111PCBA    (carte fille, suffixe préservé)
60-I778671-00  → 778671-00     (variant, suffixe préservé)

The "strip the literal 60-I prefix" rule covers all 264/264 lines. The looser "last six digits" heuristic (suggested by the customer over email) succeeds on 260 lines but corrupts the four PCBA / dash-suffix variants. We apply the strict rule. Valdosta does NOT accept customer_part_number at tier 1 — in that catalogue, customer-side aliases collide with mfr-PNs across distinct families.

4. Confidence aggregation

For a PO P, the trust score is

The router auto-approves when T ≥ 0.85.

5. Soundness

Determinism. Snake is a deterministic SAT classifier: identical input yields identical output, independent of training order. Fallback tiers are dictionary lookups and pure string functions. Output is byte-stable across plant restarts.

Auditability. Each matched line carries (method, confidence, tier, audit). For Snake predictions, the audit contains the triggered SAT clauses and the bucket depth, making every decision reconstructible by hand from data/models/<plant>/article_matcher.json.

Noise invariance. The Dana Theorem (2024) guarantees Snake's SAT formula is polynomial in |M_p| and linear in the number of layers. Adding a new SKU to the basis and retraining preserves all prior correct matches that don't collide on literal tests — in practice collisions are negligible at |M| = 9,425 (Valdosta) and |M| = 43,332 (Bordeaux).

Hallucination guard. The 0.85 floor on tier 3 is what prevents Snake from inventing plausible-looking SKUs on garbage input. Verified on Safran customer PNs (BX00201-03) and Boeing identifiers (VHP430KH3) — the matcher correctly returns ⊥.

6. Stage 2 text-mode fast path

Bordeaux's PO PDFs (Saft Ferak) carry a clean text layer that pdftotext can extract. Stage 2 first attempts a deterministic text parser; if every line item parses AND the line totals reconcile to the document total within TOTAL_RECONCILE_TOLERANCE = 1.00 EUR, we skip the VLM entirely. The reconciliation gate is the quality guarantee: arithmetic equality across 264 lines is a structural proof, not a heuristic. On Saft Ferak P2502253: 264 lines & €448,650.50 reconciled to the cent in 134 ms with zero LLM tokens for stage 2; the PO completes end-to-end in 12.8 s for $0.10 of stage-0/1/5 Haiku calls only. Glass-industry POs do NOT use this path — their layouts are too unpredictable for deterministic parsers and the VLM is load-bearing there.

7. Complexity

Per-line match: O(1) exact (indexed cascade), O(|M_p|) worst case (fuzzy tier), O(L · b) Snake. For Bordeaux's 43k-class model with L = 5, b = 50, a single prediction runs in < 5 ms on the r6i.large after the 6.2-s S3-fresh load. Pipeline overhead is dominated by the VLM stages 2 and 5 when text-mode does not engage.

8. Measured accuracy (full live sweep, 2026-05-29)

The single non-tier-1 line (AM4075307 line 2) is a 3.4% Textron customs surcharge with no Saft SKU counterpart; the 0.85 floor correctly returns ⊥ rather than hallucinate.

9. Limits & open problems

• Synonym bootstrap. Master data does not carry a native "manufacturer PN → SKU" column. Phase 2 ingests historical PO receipts to extract (mfr_pn, sku) pairs and fold them into the Snake article training set.
• Schedule-line merging. Ariba POs split a single SKU across multiple schedule lines. Stage 3 merges by identical (m, d); cross-PO merging requires ERP sync.
• Real noise injection at the leaf on the SnakeBatch v9 mesh remains unsolved (sampling from local_pop reintroduces a Bayesian contradiction); cure needs gatherer-side post-routing global injection. noise=0 is hard-coded at the leaf for now. Local Snake training on this service is unaffected.
• 1000-row PO cap. Above ~1k line items, divide-and-conquer over page chunks is required to keep stage 2 in budget. Saft Ferak's 264-line PO is well below that threshold.