Test Data Files¶

This document records the source files actually used when rebuilding the whole system from scratch: where each file comes from, where it goes, and which module it feeds. The full wipe-and-rebuild on 2026-07-14 (docker compose down -v → up -d --build → re-import every module) followed exactly this list.

fhir-code/ is not in version control at all (the files are large, some are license-restricted, and all of them can be obtained again). This list is therefore the only record of that folder — when moving to another machine, use it to restore the files. The two LOINC mapping tables the project compiled itself, which cannot be downloaded anywhere, live in data/loinc/ under version control (see below).

File list¶

These live under fhir-code/ in the repository root. Only the first 16 characters of each sha256 are listed, enough to confirm you have the same file.

File	Size	sha256 (first 16)	Module / source_role
`icd/10/icd10cm/icd10cm-table-index-2025.zip`	20 MB	`7e0fa0cae6348fb6`	icd / `icd10cm`
`icd/10/icd10pcs/icd10pcs_tables_2025.zip`	648 KB	`728d9b8f5c315548`	icd / `icd10pcs`
`icd/10/1.2023年中文版ICD-10-CM_PCS_1131118V3(…).xlsx`	7 MB	`1bf34d1f92930f29`	icd / `icd_zh_tw`
`loinc/2.80/Loinc_2.80.zip`	74 MB	`5a6b10dd6ed29704`	loinc / `loinc`
`snomed/SnomedCT_InternationalRF2_PRODUCTION_20250601T120000Z.zip`	540 MB	`3d6c189288e375d4`	snomed / `snomed_ct`
`rxnorm/RxNorm_full_06032024.zip`	241 MB	`fc612ca0abee8955`	rxnorm / `rxnorm_full`
`twcoreig/v1.0.0/package.tgz`	2.8 MB	`2fa1419fa48d1545`	ig / `ig`
(only after download) `36_2.csv`	44 MB	—	drug / `drug_index_csv`
(recommended since 2026-10) `icd/10/icd10cm/april-1-2026-code-descriptions-tabular-order.zip`	—	—	icd / `icd10cm`
(optional, generated by a script) `icd10who.json`	—	—	icd / `icd10_who`

ICD-10-CM now uses the CMS order file

icd10cm-table-index-2025.zip (tabular XML) in the table above is the file the 2026-07-14 rebuild used. Since 2026-10 the CMS "Code Descriptions in Tabular Order" zip (2026-04-01 release, containing icd10cm_order_2026.txt, 98,186 codes) is recommended instead: only this complete edition (icd.editions.complete = true) makes the FHIR validator report unknown ICD-10-CM codes as errors. The WHO ICD-10 icd10who.json is fetched from tx.fhir.org with node-server/scripts/validator-audit/harness/fetch_tx_terms.py. Neither is in the local fhir-code/ yet, hence no sha256.

The local loader (without an admin upload) looks for the ICD-10-CM zip in fhir-code/icd/10/icd10cm/ in this order: the ICD_CM_ZIP env var → a CMS order-file zip (name contains order / code-descriptions) → a zip whose name contains table-index / tabular (last by name within each kind) → the legacy default icd10cm-table-index-2025.zip. So dropping the order-file zip into that folder is enough; the old table-index zip can stay.

umls/umls-2024AA-metathesaurus-full.zip (4 GB) is also in the folder, but no import flow currently uses it — skip it when rebuilding.

The two LOINC mapping tables kept in version control¶

data/loinc/taiwan_mapping.csv (loinc / loinc_taiwan_mapping) and data/loinc/lab_reference_ranges.csv (loinc / loinc_reference_ranges) were compiled by the project itself (Taiwan lab code mappings and reference values collated from several medical centres) and cannot be downloaded anywhere, so they live under data/ and travel with the repo — not under fhir-code/ (that entire folder stays out of version control). Both roles are optional in the source catalogue, but without them there are no Chinese names or reference ranges (interpret_lab_result cannot interpret anything) — remember to upload them alongside the LOINC zip.

Where each file comes from¶

ICD-10-CM / PCS: the official CMS annual release (for CM, the 2026-04-01 "Code Descriptions in Tabular Order" is recommended) (https://www.cms.gov/medicare/coding-billing/icd-10-codes). The Chinese-name XLSX comes from the MoHW Chinese edition of the ICD-10-CM/PCS mapping file.

LOINC 2.80: the Regenstrief Institute (https://loinc.org/downloads/); requires registering an account to download.

SNOMED CT International RF2: requires an active license (obtainable through an NLM / UMLS account); https://www.nlm.nih.gov/healthit/snomedct/. Not redistributable.

RxNorm Full Release: NLM, requires a UMLS account; https://www.nlm.nih.gov/research/umls/rxnorm/. This project uses it only as concept-only reference terminology (for IG ValueSet expansion), not as a standalone drug tool.

TW Core IG package.tgz: obtain tw.gov.mohw.twcore from https://packages.fhir.org. The import fetches dependency packages recursively and automatically (hl7.terminology, hl7.fhir.r4.core, hl7.fhir.r4.examples, and so on, from packages.fhir.org with packages2.fhir.org as fallback); the 2026-07-14 rebuild ended with 9 packages and roughly 21,000 artifacts in the database — so this one tgz is all you need to prepare. Packages fetched as dependencies are marked install_origin = 'dependency'.

TFDA drug licenses 36_2.csv: not in the repo, and there is no need to place it in fhir-code/ by hand. Download it directly:

curl -L "https://data.fda.gov.tw/data/opendata/export/36/csv" -o drug36.zip
# This is a zip containing 36_2.csv (roughly 44 MB, 71,921 rows, UTF-8 with BOM)

The two other plausible-looking URLs (cacheData/36_2.csv and codedata/datadownload/36) both return 404.

Health supplements and food nutrition need no source file: health_supplements_sync / food_nutrition_sync call the TFDA Open Data API directly.

Rebuild procedure¶

Restore settings (OCR / TFDA / registry addresses, model provider accounts and routes, passkeys) from a settings export (admin console System → Backups & settings, or the API below; the export format is currently v3, and v1 / v2 files still import):

curl -X POST localhost:8080/admin/api/settings/import \
     -H 'Content-Type: application/json' \
     --data-binary @tw-health-settings-YYYY-MM-DD.json

That export contains API keys and passkeys — keep it out of version control. Keep your own backup.

Upload source files on each module page's "Data lifecycle" tab (or through the API; note the Content-Type: application/octet-stream header, without which express's 4 MB parser intercepts the body):

curl -X POST "localhost:8080/admin/api/uploads?module_key=icd&source_role=icd10cm&filename=april-1-2026-code-descriptions-tabular-order.zip&auto_activate=true" \
     -H 'Content-Type: application/octet-stream' \
     --data-binary @fhir-code/icd/10/icd10cm/april-1-2026-code-descriptions-tabular-order.zip

Then queue the jobs in order: icd_import, loinc_import, snomed_import, rxnorm_import, ig_import (which needs job_options.object_key, the MinIO object key of that source file), health_supplements_sync, food_nutrition_sync, and drug_index_import.

The drug pipeline chains automatically

A successful drug_index_import auto-chains into drug_enrichment, which then chains into drug_analysis. Enrichment crawls the official TFDA site and analysis calls OCR + an LLM — both cost money and time, and there are over twenty thousand licenses pending. Auto-chaining is now capped at 200 per batch (DRUG_AUTOCHAIN_BATCH_LIMIT) and advances gradually; stop it from the Jobs page if you do not want it to run. For a small-scale check, queue a drug_enrichment with {"limit": 30} yourself.

The data volumes this list produces¶

The actual result after the 2026-07-14 wipe-and-rebuild, usable for checking whether an import completed (the ICD diagnosis count is from the tabular XML used then; with the 2026-04-01 order file it is 98,186 codes):

Module	Rows
ICD	46,498 diagnoses + 78,948 procedures
LOINC	104,672 concepts
SNOMED CT	373,972 concepts
RxNorm	222,199 concepts
FHIR IG	9 packages / 20,996 artifacts
Drugs	66,395 licenses
Health supplements	565
Food nutrition	1,702

Semantic search needs the *_embed jobs run separately (which need a reachable embedding endpoint, configured under Integrations → Model endpoints, e.g. Ollama). Without embeddings, Chinese keyword search finds almost nothing — Postgres's simple tokenizer treats a whole Chinese product name as a single token, so substring matching fails. This is not a bug; the design relies on vector search to cover it.