Paperless-ngx Taxonomy Rebuild Proposal

Generated 2026-02-15 — Review and mark up before CoS executes

Current State

3,100 documents total. 2,800 (90%) have no document type and no correspondent assigned.

19 document types (some duplicates), 80 tags (doing 6 different jobs), 55 correspondents (mostly clean), 0 workflows, 0 custom fields, 0 saved views.

The ML classifier is trained but returns None for everything — not enough training examples per category.

Proposed Document Types (12)

Document types answer: "What kind of document is this?"

New Type	Replaces	Match Pattern (suggested)	Notes
Bank Statement	Banking (1 doc)	statement, account summary	Monthly/quarterly statements from any bank
Bill	Bill (2 docs)	amount due, balance due, pay by	Utility, service, any invoice to pay
Correspondence	Correspondence (6 docs)		Letters, notices, general mail
Deed / Title	Heron House (9), Sterling House (3)	deed, title, conveyance, grantor	Property deeds, vehicle titles, legal ownership docs
Earnings Statement	Earnings Statement (60 docs)	earnings statement, pay stub, net pay, gross pay	Pay stubs, W-2s, 1099s
Estimation / Quote	Estimation (1 doc)	estimate, quote, proposal	Contractor estimates, repair quotes
Insurance Document	—	policy, premium, coverage, declaration	Policies, declarations, claims — new type
Medical Bill	Medical Bill (45) + Medical-Bill (9)	patient responsibility, amount due, medical	Merge the duplicates
Medical Record	Medical Record (2) + Medical (1) + Medical-Back (6)	diagnosis, patient, treatment plan, lab results	All clinical records, lab results, treatment plans
Mortgage Document	Mortgage Document (1 doc)	mortgage, escrow, loan	Mortgage statements, escrow analyses, refinance docs
Receipt	Receipt (1 doc)	receipt, paid, transaction	Proof of payment
Tax Document	Taxes (23) + Property Taxes (1)	tax return, form 1040, property tax, assessed value	Federal, state, property taxes

Removed: "Heron House" and "Sterling House" are properties, not document types — they move to the Property custom field. "Manual" (1 doc) — owner's manuals probably don't need to be in Paperless. "Payment" (3 docs) — merge into Receipt or Bill depending on context. "Personal Record" (6 docs) — too vague, reclassify into specific types.

Proposed Tags (~18 category tags)

Tags answer: "What topic/category does this fall under?" — only one job.

Tag	Replaces	Notes
Banking	Bank (46), MidWestOne (45), USBank (1), Statement (27)	All banking-related. Correspondent identifies which bank.
Dental	Insurance-Dental (0 docs)	Dental-specific medical/insurance
Employment	Payroll (90), Income (51), Certification (1), Benifits (3)	Anything job/career related
Insurance	Insurance (44), Insurance-Auto (7), Insurance-Home (5), Insurance-Health (2), Insurance-Motorcycle (2), Auto-Owners (1)	Sub-types not needed — document type + correspondent tells you what kind
Legal	TitleInsurance (1)	Legal documents, contracts, agreements
Medical	Medical (88), Laboratory (26), Endoscopy (1), GeneralVisit (0), SuntreeInternalMedicine (0)	All health/medical. Doctor is the correspondent.
Military / VA	VA_Claim (17), VA (2), USMC (8)	All military service and VA claims
Mortgage	Mortgage (1), 2020 ReFi (3), 2025 ReFi (1)	Refi events don't need their own tag
Property	Property (4), Taxes-Property (1), Home (44)	Property-related (which property is the custom field)
School / Education	School (3)	Education records, transcripts, student loans
Taxes	Taxes (40), Taxes-2016 thru Taxes-2024 (all), W-2 (1), 1099 (1)	Year is already in created_date. W-2/1099 are document sub-types.
Vehicle	Registration (2)	Vehicle titles, registration, insurance
Social Security	Social Security (3)	SSA correspondence, benefits
ToDo	ToDo (392) [INBOX]	Keep as inbox tag — working as intended

Tags being eliminated (moved to other dimensions):

Year tags (1989, 1999, 2003...2025) → Already captured by created_date
People tags (Philip, Rosanne, Dusty, Krystal, Frankie, Willa) → Custom field "Family Member" or correspondents
Location tags (Heron, Sterling Dr, GlenGardner, Flintlock, Ventura) → Custom field "Property"
Correspondent-as-tag (MidWestOne, Auto-Owners, USBank) → Already the correspondent field
Specific procedures (Endoscopy, GeneralVisit) → Too granular for tags; that detail is in the document content
Misc: Pending (1), Personal (1), Payment (3), DormantAccount (1), Estimate (0), MarriageCertificate (0), Gomez Justin (1) → reclassified or deleted

Proposed Custom Fields (new)

Field Name	Type	Values	Purpose
Family Member	Select	Philip, Rosanne, Dusty, Krystal, Frankie, Rick, Marjorie, Skyler, Ava, Emery, Ramona, Lourdes	Who in the family this document pertains to
Property	Select	Heron, Sterling Dr, GlenGardner, Flintlock, Ventura	Which property (if applicable)

Paperless custom fields with type "Select" were added in v2.x. Your instance should support this. If not, these could be implemented as a small dedicated tag group with a color convention instead.

Correspondent Cleanup

Action	Correspondent	Reason
Delete	Estimate	Not a correspondent (0 docs)
Delete	Phil	Not a correspondent (0 docs) — that's you
Merge	SF Insurance → Security First Insurance	Same company, 1 doc
Add	Willa Price, NP	Currently a tag, should be correspondent
Add	Dr. Justin Gomez	Currently a tag ("Gomez, Justin"), hepatologist
Keep all others	(53 correspondents)	These are clean and correct

Pipeline Plan

Phase 1: Taxonomy Rebuild (CoS does this via API)

Create new document types, delete/merge old ones
Create custom fields (Family Member, Property)
Delete year/people/location/duplicate tags
Create clean category tags with match patterns
Clean up correspondents
Re-assign the ~181 already-typed docs to new types
Set up match patterns on correspondents and document types

Phase 2: Backlog Classification (Ollama pipeline — Phil runs)

Script fetches document IDs from Paperless API
For each unclassified doc: fetch OCR text, send to Ollama, get classification
Ollama returns: document type, correspondent, tags, family member, property, suggested title
Script updates Paperless via API, moves ID to "processed" list
Resumable — crash-safe checkpoint file
Metadata saved to JSONL with redacted PII (initials, last-4 account numbers)

Phase 3: Workflows (CoS sets up after backlog is done)

Auto-assign correspondent when content matches known patterns
Auto-assign document type based on content keywords
Auto-remove ToDo inbox tag when document has been classified
New scans auto-classify going forward

Phase 4: Tika/Gotenberg (future)

Add containers to docker-compose for Office doc support
Remove Office file types from ignore patterns
Re-ingest any XLS/DOC/DOCX files from backups if needed

Redaction Rules (for metadata JSONL output)

Real Name / Data	Redacted As
Philip / Phil Langeberg	P.L.
Rosanne Langeberg	R.L.
Dusty Langeberg	D.L.
Krystal Langeberg / Kirkland	K.L.
Frankie / Francine / Frank Dupre / Langeberg	F.L.
Rick / Ricky Langeberg	Ri.L.
Marjorie / Marge Zetocha / Job / Morrison	M.Z.
Skyler Snyder	S.S.
Ava Kirkland	A.K.
Emery Kirkland	E.K.
Ramona Acevedo-Torrellues	R.A.
Lourdes / Kitty Taylor / Acevedo	L.A.
SSNs	xxx-xx-1234
Account numbers	****1234
Addresses	[ADDR-HOME] or [ADDR-OTHER]
Phone numbers	(xxx) xxx-1234