Skip to main content
Health data arrives messy — “血糖(空腹)”, “FBG”, “Glucose, fasting” are the same measurement in three costumes, with units to match. Mirobody’s data plane standardizes every write so that the agent (and your queries) see one coherent, coded dataset. This page explains the pipeline; you never call it directly — it runs inside POST /v1/data, POST /v1/files, and POST /v1/extract.

​
The pipeline

document / record
   │
   ├─ 1. OCR & text extraction        PDF / image → text; Excel → structured rows
   │
   ├─ 2. LLM reading extraction       schema-constrained: {indicator, value, unit, measured_at}
   │                                  (documents only — structured /v1/data records skip this)
   ├─ 3. LOINC resolution             deterministic name → code: curated corpus + embedding
   │                                  retrieval. The LLM is NEVER asked to guess a code —
   │                                  no confident match means loinc_code = null, not a wrong code.
   ├─ 4. UCUM unit normalization      "mg/dl" → mg/dL, value parsed to a number
   │
   └─ 5. FHIR mirror                  a FHIR R4 Observation per reading (fhir_resource_id)
The critical design choice is step 3: codes are resolved deterministically, not generated. Free-text indicator names are matched against a curated LOINC corpus via embedding retrieval with a confidence threshold. LLMs are great at reading documents (step 2) and terrible at reciting code systems — a hallucinated LOINC code is worse than none, so an unconfident match returns null and keeps the raw name.

​
Before / after

What you send (or what a lab PDF says) vs. what the store holds:
Before (as written)After (standardized)
Indicator"血糖(空腹)" / "FBG" / "Glucose, fasting"loinc_code: "1558-6", canonical_name: "Fasting glucose [Moles/volume] in Serum or Plasma"
Value"5.4 mmol/L" (one string)parsed_value: 5.4, parsed_unit: "mmol/L" (UCUM)
Unit spelling"mg/dl", "MG/DL", "mg/dL""mg/dL" (one UCUM form)
Interopfree textFHIR R4 Observation (fhir_resource_id)
Original—kept: value stays the human-readable string; raw names/units preserved
Three spellings of the same test become one series — trend queries, the agent’s data tools, and your own analytics all see it as one indicator.

​
Where you see it

SurfaceWhat surfaces
POST /v1/dataResponse counts standardized alongside ingested.
GET /v1/dataEvery row carries parsed_value / parsed_unit / loinc_code / canonical_name / fhir_resource_id.
POST /v1/extractThe whole pipeline as a synchronous call — dry-run by default, so you can inspect standardization before committing a write.
The agentGrounded answers query the standardized series — which is why “how’s my glucose?” finds records written as “FBG”.

​
Cookbook: one call, report → structured data

# Inspect first (nothing persisted) …
curl https://test-mirobody-api.thetahealth.ai/v1/extract \
  -H "Authorization: Bearer $MIROBODY_API_KEY" \
  -F "user=alice" -F "file=@lab_report.pdf"

# … then commit the same extraction
curl https://test-mirobody-api.thetahealth.ai/v1/extract \
  -H "Authorization: Bearer $MIROBODY_API_KEY" \
  -F "user=alice" -F "file=@lab_report.pdf" \
  -F "store=true" -F "retention=permanent"
See Extract for the full row shape (indicator_raw, loinc_code, confidence, …).