Tools that can split mixed shipping packets (BOL + rate con + POD) and return separate structured outputs per document type

Most freight and logistics teams don’t struggle with “reading” a Bill of Lading. They struggle with the 40‑page mixed packet: BOL + rate confirmation + POD + random emails + certs, all in one PDF or email thread, from thousands of carriers. You don’t need another OCR demo; you need tools that can split mixed shipping packets reliably and return separate, schema‑enforced outputs for each document type.

Quick Answer: The tools that actually work on mixed shipping packets combine three capabilities: semantic document splitting, per‑type classification (BOL vs rate con vs POD vs “other”), and extraction into strict, per‑document schemas. Platforms like Bem implement this as deterministic workflows—split → classify → extract → enrich → validate—so each packet call returns multiple structured outputs (or flagged exceptions) instead of a single blob of text.

Why This Matters

If your TMS, billing engine, or claims workflows depend on BOLs, rate cons, and PODs, mixed packets are the bottleneck. Per‑page OCR and generic “document AI” can show you text and bounding boxes, but they don’t cleanly separate document types or normalize data into something your ERP can trust.

The impact is operational, not academic:

• Slow carrier onboarding.
• Missed accessorials and mis‑rated freight.
• Manual triage of edge cases and weird layouts.
• Claims and payment disputes because totals don’t reconcile.

Tools that can split mixed shipping packets and return separate structured outputs per document type let you ship workflows in hours instead of months—and keep them running when a carrier changes their layout or merges three PDFs into one.

Key Benefits:

• Deterministic multi‑doc handling: One packet in, multiple validated JSON outputs out—BOL, rate con, POD—each with its own schema and confidence.
• Less manual triage: Low‑confidence fields and odd packets are automatically routed to review instead of silently corrupting your TMS.
• Faster change management: Versioned functions/workflows and regression tests let you update extraction logic per document type without breaking production.

Core Concepts & Key Points

Concept	Definition	Why it's important
Packet splitting	Automatically detecting document boundaries inside a mixed PDF/email and splitting into logical documents (BOL, rate con, POD, etc.).	Without robust splitting, everything downstream is fragile—wrong pages end up in the wrong schema or get double‑counted.
Document classification	Semantic identification of each split segment’s type and language (e.g., “Bill of Lading (EN)”, “Rate Confirmation (ES)”).	Enables routing to the correct extraction schema and workflow (AP, claims, disputes) with different field sets and business rules.
Schema‑enforced extraction	Converting each identified document into JSON that must conform to a pre‑defined schema, with per‑field confidence and hallucination checks.	Turns messy packets into reliable, machine‑readable data that your TMS, billing system, or data warehouse can trust—or explicitly flags exceptions.

How It Works (Step‑by‑Step)

Below is how a production‑grade tool like Bem handles mixed shipping packets end‑to‑end.

1. Ingest & Split the Packet

   • Input: a single PDF or email attachment containing BOL + rate confirmation + POD (and often more: invoices, customs docs, emails).
   • The system runs a Split function that uses layout signals, semantics, and prior patterns to identify document boundaries.
   • Output: discrete document slices with page ranges, e.g.:
     • Document 1: pages 1–3 → candidate “Bill of Lading”
     • Document 2: pages 4–5 → candidate “Rate Confirmation”
     • Document 3: page 6 → candidate “Proof of Delivery”

2. Classify Each Document Type

   • Each slice is passed through a Route / Identify function.
   • The tool assigns a type and subtype, e.g.:
     • type: "bill_of_lading", locale: "en-US"
     • type: "rate_confirmation", lane_type: "LTL"
     • type: "proof_of_delivery", mode: "truckload"
   • Misclassifications are caught by confidence thresholds; low‑confidence docs route to an exception queue or generic “unknown_shipping_doc” schema instead of being forced.

3. Extract into Schema‑Enforced JSON

   • For each recognized type, the system applies a Transform / Extract function bound to a strict schema. For example:

   Bill of Lading schema (simplified):

   {
     "bol_number": "string",
     "shipper_name": "string",
     "consignee_name": "string",
     "carrier_name": "string",
     "origin_address": "string",
     "destination_address": "string",
     "container_numbers": ["string"],
     "seal_numbers": ["string"],
     "gross_weight_kg": "number",
     "line_items": [
       {
         "sku": "string",
         "description": "string",
         "weight_kg": "number",
         "quantity": "number"
       }
     ],
     "eta": "string"
   }
   

   Rate confirmation schema (simplified):

   {
     "rate_con_number": "string",
     "carrier_name": "string",
     "bol_number": "string",
     "lane": {
       "origin": "string",
       "destination": "string"
     },
     "base_rate": "number",
     "fuel_surcharge": "number",
     "accessorials": [
       {
         "type": "enum",
         "amount": "number"
       }
     ],
     "currency": "string"
   }
   

   POD schema (simplified):

   {
     "bol_number": "string",
     "delivery_date": "string",
     "received_by": "string",
     "signature_present": "boolean",
     "exceptions": ["string"],
     "pages": "number"
   }
   

   • Extraction is layout‑aware (tables, line items) and enforced by architecture: the workflow either returns schema‑valid JSON with per‑field confidence, or it flags the document as an exception. No silent partials.

4. Enrich with Internal Data

   • A production tool should support Enrich/Hydrate steps against your own Collections:
     • Match carrier_name to a carrier master list with a match score.
     • Map origin/destination to internal location IDs.
     • Attach GL codes or cost centers based on customer or lane.
   • These enrichment steps run after extraction and are versioned independently.

5. Validate, Join, and Sync

   • Validate: enforce cross‑doc rules—e.g., the BOL gross weight must equal the rate con billing weight; BOL number must match across BOL, rate con, and POD.
   • Join: create a normalized “shipment” entity by joining BOL + rate con + POD outputs under a single shipment ID.
   • Sync: send results to downstream systems (TMS, billing, claims) via REST callback, webhook, or queue.

6. Observe, Evaluate, and Improve

   • Production pipelines log per‑field confidence, hallucination flags, and evaluation scores (F1, pass rate) against golden datasets.
   • Versioning and rollback on each function/workflow let you iterate on extraction schemas or models without breaking existing packets.
   • Idempotency keys ensure re‑processing a packet won’t double‑post to your ERP.

Common Mistakes to Avoid

• Treating per‑page OCR as the whole solution:
  How to avoid it: Don’t stop at “we can read the text.” Demand packet‑level outcomes: “For this PDF, give me a validated BOL object, a rate confirmation object, and a POD object—or tell me exactly why you can’t.”

• No schema or validation layer:
  How to avoid it: Define strict, versioned schemas per document type and enforce them. If a tool can return arbitrary JSON blobs with missing fields and no exception handling, you’ll end up debugging production invoices by hand.

• Ignoring edge cases and packet noise:
  How to avoid it: Plan for emails, cover pages, carrier marketing pages, and legacy scans. Your workflow should explicitly route “non‑shipping docs” or “unknown” instead of trying to jam everything into BOL/rate con/POD schemas.

• Hard‑coding layouts instead of using semantic classification:
  How to avoid it: Choose tools that classify based on semantics and patterns, not just fixed templates. Carriers change PDFs. Your pipeline shouldn’t need a sprint every time they move a logo.

Real‑World Example

A logistics team receives a 4‑page international shipping packet as a single PDF:

1. Packet Content:

   • Page 1: Bill of Lading
   • Page 2: Customs Declaration
   • Page 3: Payment Receipt
   • Page 4: Release Order

   In a domestic context, swap these for: BOL, rate confirmation, POD, and an invoice, all attached to the same email.

2. What a Bem‑style pipeline does:

   • Split: A split-documents_* function identifies four logical documents with ~99% confidence and assigns page ranges.
   • Classify: Each document is classified—BOL, customs declaration, payment receipt, release order (or for domestic freight: BOL, rate con, POD, invoice).
   • Extract:
     • The BOL is parsed into structured fields: shipper, consignee, vessel/truck, ports, container numbers, seal numbers, gross weight, line items—each with a confidence score.
     • The rate con is parsed into lane, base rate, fuel, accessorials, and currency.
     • The POD is parsed into delivery date, receiver, exceptions, and signature status.
   • Enrich & Validate:
     • Carrier names are matched against your internal carrier master (with match confidence).
     • BOL numbers are cross‑checked across BOL, rate con, and POD; discrepancies are flagged.
   • Sync:
     • A clean “shipment” object is posted to your TMS, and journals are created in your financial system automatically if all validations pass.
     • If the POD’s delivery date conflicts with the TMS, the packet is routed to a review queue, not silently accepted.

3. Operational outcome:

   • No human ever manually splits the PDF.
   • No one copies totals into a spreadsheet to check them.
   • Exceptions are explicit, auditable, and queued for operators.
   • You can run F1‑style evals on BOL/RateCon/POD extraction quality and roll back changes if a new version regresses.

Pro Tip: When evaluating tools, send your ugliest mixed packets—the ones with carrier logos, fax artifacts, and mixed language pages—and measure not just field‑level accuracy, but packet‑level behavior: Did it split correctly? Did it identify each document type? Did it enforce your schemas and business rules?

Summary

Tools that can split mixed shipping packets and return separate structured outputs per document type don’t just “run OCR.” They:

• Split mixed PDFs/emails into logical documents.
• Classify each slice as BOL, rate con, POD, invoice, or other.
• Extract into strict, per‑type JSON schemas with confidence and hallucination detection.
• Enrich, validate, and route results into your TMS, billing, and claims systems with versioned, idempotent workflows.

This is the difference between a demo parser and a production layer for unstructured freight data. If you’re still hand‑splitting packets or reconciling totals in Excel, you’re not using the right tools.

Next Step

Get Started