yesil-health-ai-api (Python)

Official Python client for the Yesil Health AI enterprise (B2B) API.

Synchronous, built on httpx. The hard part of integrating — parsing the typed Server-Sent Events stream — is handled for you. Mirror of the TypeScript client @yesilsci/health-ai-api.

Install

pip install yesil-health-ai-api

Quick start

import uuid
from yesil_health import YesilHealthClient

client = YesilHealthClient(
    base_url="https://api.yesilhealth.com",
    api_key="...",  # sent as X-API-Key
)

for ev in client.chat_stream({
    "question": "Is metformin safe for a patient with stage 3 CKD?",
    "request_id": str(uuid.uuid4()),       # idempotent billing — recommended
    "demographics": {"age": 62, "sex": "female"},
    "health_context": "eGFR 45 mL/min/1.73m². On lisinopril 10mg daily.",
}):
    if ev["type"] == "delta":
        print(ev["content"], end="", flush=True)

One-shot (no live rendering):

result = client.chat({"question": "..."})
print(result["text"])

The client is a context manager:

with YesilHealthClient(base_url=..., api_key=...) as client:
    ...

Authentication

Every request is authenticated with your enterprise API key via the X-API-Key header. Keep it server-side; never ship it in a client app.

The event stream

chat_stream() yields event dicts. Route on event["type"]:

Forward compatibility. The API only ever adds fields and event types within a contract version. Always ignore unknown type values — this client yields them as plain dicts rather than raising.

Errors

Non-2xx responses raise YesilApiError with status, code (server error_code), and message:

from yesil_health import YesilApiError

try:
    client.chat({"question": "..."})
except YesilApiError as e:
    if e.code == "RATE_LIMITED":  # 429
        ...
    if e.status == 402:           # quota exhausted / billing
        ...

Evidence search

Ranked literature retrieval for a query, no LLM answer synthesis. Requires the evidence:read scope on your API key.

res = client.evidence_search({
    "question": "omega-3 and cardiovascular risk reduction",
    "needs_trials": False,
})
print(res["papers"])

Account endpoints

client.billing()  # billing status
client.quota()    # quota / rate-limit status
client.usage()    # usage report

Versioning

Defaults to contract v1 (/api/v1). For V2 chat, change only chat_endpoint; a global version switch also changes billing, evidence and PDF paths. V1 remains available. V2 access must be enabled for the tenant by Yesil Health. Kept in lockstep with the TypeScript client — see backend/sdk/CONTRACT_SYNC.md.

V2 engine with the V1 chat format (local SDK addition)

Pass chat_endpoint="/api/v2/chat/stream/compat" to the client constructor and leave the global API version at its default V1. Only chat uses the adapter; billing/evidence/documents keep their existing paths. Existing delta/citation/memory handlers work unchanged. The adapter requires tenant V2 access. Enabled reminder/measurement workers emit additive reminder_action / measurements_extracted events; apply their action once per operation_id. Compat also emits handoff_recommendation and service_recommendation when a validated link or approved catalog selection is ready. Markdown deltas still contain a link fallback, so avoid rendering the same call-to-action twice. Old readers may ignore unknown events. V2 memory is institution-controlled; extract_memory is accepted but ignored.

For native V2 use chat_endpoint="/api/v2/chat/stream". The SDK exposes state_patch, answer and normalized card event types. The answer event is a final snapshot of the text from deltas; replace the current message with it if needed instead of appending it as a second message.

Both V2 chat endpoints return additive done.billing and done.usage fields. Billing lines show the tenant's configured rate, discount, charged amount, and the worker/result code that triggered an add-on. Worker summaries contain only module names, status/outcome codes, durations, and result counts; they do not contain patient text or source titles. Standard research and reminders are part of chat. Memory and measurement share one add-on line, as do institution search, approved service matching, and handoff. Deep research and useful structured PDF results have separate lines. Exact rates come from the tenant's rate card and are not hardcoded by this client.

V2 document processing is selected independently, leaving billing/evidence on V1:

v2 = YesilHealthClient(
    base_url="https://api.yesilhealth.com",
    api_key="...",
    chat_endpoint="/api/v2/chat/stream",
    document_endpoint="/api/v2/documents/process",
)
parsed = v2.parse_document(open("lab.pdf", "rb"), document_hint="auto")
# Keep parsed["context_block"] in your own patient context and send it as health_context.

The stateless institution loop is documented in the console API guide (Conversation state): persist memory signals and proposed actions, then send the institution's current reminders and relevant context on later turns. PDF billing.settlement_status is settled, trial, or not_chargeable; empty/unusable parse results do not receive a document charge. Version 0.5.0 adds the V2 endpoint options. Full reference: the API guide in the Yesil Health console (API Docs).