haipipe-individual-inference

name	haipipe-individual-inference
description	Per-individual endpoint inference test: loads one individual from _WorkSpace/A-User-Store, builds an Endpoint_Set dataframe_records payload, POSTs it to a deployed endpoint (local FastAPI / Databricks / SageMaker -- same wire contract), and prints the forecast. Use to smoke-test that a deployed endpoint accepts individual data. Trigger: individual inference, test endpoint with individual, /haipipe-individual-inference.
argument-hint	--individual <id_or_path> [--endpoint-url URL] [--json]
allowed-tools	Bash, Read
metadata	{"version":"0.1.1","last_updated":"2026-07-04","summary":"Per-individual endpoint inference test."}

Skill: haipipe-individual-inference

Single-individual inference smoke test against a deployed Endpoint_Set.

Pipeline:

  Subject-XX/                                            (haipipe-individual layout)
    ├── 1-SourceStore/{CGM,Diet,Ptt,...}.parquet
    └── manifest.yaml
            │
            ▼  load_patient_ctx()
       patient_ctx dict
            │
            ▼  build_payload()                 (Endpoint_Set wire contract)
       {dataframe_records: [{TriggerName_to_CaseTriggerList, inference_form: {ElogBGEntry, Patient}}]}
            │
            ▼  call_predict()  ── HTTP POST ──► <endpoint_url>/invocations
       forecast JSON

The wire payload matches the Endpoint_Set's documented Input2SrcFn contract (Format 1 — dataframe_records). Same payload Databricks Model Serving and SageMaker would consume; only endpoint_url differs.

---

Layout

src/
  load_patient.py    Subject-XX → patient_ctx dict (parquet → DataFrames)
  build_payload.py   patient_ctx → dataframe_records JSON (Endpoint_Set contract)
  client.py          POST dataframe_records → forecast (auth + retry-friendly)

scripts/
  show_ctx_cli.py            inspect an individual's loaded context
  test_individual_predict.py    end-to-end: individual id → POST → forecast

---

Quickstart

Spin up a local endpoint server (skill: haipipe-end-deploy-local):

ENDPOINT_PATH=_WorkSpace/6-EndpointStore/<endpoint_name> \
    python Tools/plugins/haipipe-toolkit/skills/task/3_end/haipipe-end-deploy-local/scripts/serve_local.py

In another shell, hit it with an individual:

python Tools/plugins/haipipe-toolkit/skills/task/4_individual/haipipe-individual-inference/scripts/test_individual_predict.py \
    --individual Subject-18

Override the URL to test against a Databricks endpoint:

CGM_ENDPOINT_URL=https://<workspace>/serving-endpoints/<name>/invocations \
    python ...test_individual_predict.py --individual Subject-18

---

Individual resolution

--individual accepts:

Form	Example
absolute path	/home/.../UserGroup-WellDoc2022CGM/Subject-18
UserGroup-X/Subject-Y	UserGroup-WellDoc2022CGM/Subject-18
Subject-Y (auto-resolved if unique)	Subject-18

Ambiguous bare ids raise ValueError listing the candidates.

---

Wire contract

{
  "models": ["endpoint_cgm_patchtst_ohio/v0001"],
  "dataframe_records": [{
    "TriggerName_to_CaseTriggerList": {
      "CGM5MinEntry": [{"PatientID": "...", "ObsDT_UTC": "...", "TimezoneOffset": -240}]
    },
    "inference_form": {
      "ElogBGEntry": {"PatientID": [...], "ObservationDateTime": [...], "BGValue": [...]},
      "Patient":     { ...  ...  ...
         
    
  

NaN / NaT in source parquet → null in JSON (handled by _df_to_columnar).

---

Failure modes

Symptom	Likely cause	Fix
500 Inference error: 'TriggerName'	Endpoint is missing prefn_config.json	Regenerate the Endpoint_Set with PreFnPipeline artifacts
No model to call in current ModelSeries	Payload models field doesn't match endpoint's External_to_Local_ModelSeries	Check meta.json and pass the right model id
ValueError: Cannot resolve individual	--individual shorthand matched 0 or >1 paths	Use full path or UserGroup-X/Subject-Y
ValueError: No CGM data	Individual parquet empty/missing	Check 1-SourceStore/CGM.parquet exists and has rows

---

Reuses

• Loads the individual layout produced by haipipe-individual (sibling skill).
• Targets endpoints produced by haipipe-end-endpointset and served by haipipe-end-deploy-{local,databricks,sagemaker}.
• Used by agent projects (e.g. agent-cgm) as the patient-data loader and payload builder for their LangGraph nodes.