Live models

Evidence Q&A

Model: aug-evidence-qa · Status: Preview · Scope: agents:evidence_qa · Sessions: no

Answers clinical questions with every clinical sentence cited to the evidence behind it, and the certainty of that evidence stated. Built for clinicians, medical educators and clinical content teams. Answers stream as they are written.

Request

JSON

{
  "model": "aug-evidence-qa",
  "stream": true,
  "messages": [
    {"role": "user", "content": "Can metformin be continued in a patient with eGFR 25, and what are the risks?"}
  ]
}

The last user message is the question (up to about 2,000 tokens including earlier turns).
Earlier user / assistant turns (up to four) are used to understand follow-ups such as "and in pregnancy?". The answer itself is always grounded afresh.
The model has no session; send the conversation each time, as with any chat model.

Response

The answer is in message.content (or the streamed delta.content):

Text

Metformin should **not** be continued in a patient with an eGFR of 25 mL/min/1.73 m² [5][9]. It is
contraindicated in severe renal impairment, defined as an eGFR below 30 mL/min/1.73 m² [5][9][23].

**Risks**
* **Lactic acidosis:** the main risk of continuing metformin in renal impairment [21][24]; very
  low-certainty evidence describes it as rare but potentially fatal [21].
* **Vitamin B12 deficiency** with long-term use [6][20].

Citations arrive with the final message or chunk, in augsidius.citations:

JSON

[
  {"index": 5, "claim": "Metformin is contraindicated in Severe renal impairment (eGFR < 30 mL/min/1.73 m²)",
   "certainty": "moderate", "detail": null,
   "sources": [{"type": "pubmed", "pmid": "32138474", "title": "…", "year": 2020,
                "url": "https://pubmed.ncbi.nlm.nih.gov/32138474/", "quote": "…"}]},
  {"index": 21, "claim": "Metformin causes Lactic Acidosis", "certainty": "very_low", "…": "…"}
]

Field	Description
`index`	The number used in `[n]` markers. Only cited statements are listed; numbering can have gaps.
`claim`, `certainty`, `detail`, `sources`	See Evidence.

message.annotations holds the same links as OpenAI url_citation annotations (character offsets of each [n] marker), so OpenAI-aware chat UIs render them without extra work.

Showing answers well

Render [n] as superscript links that open the citation: claim, certainty and the source with its quote.
Show certainty next to claims where it is low, very_low or ungraded.
Keep the question and answer together in your audit log with x-request-id.

What the model will and will not do

Only answers from the evidence. If the evidence does not cover the question (or part of it), the answer says so instead of filling the gap. Such answers come back with no citations.
Clinical questions only. Conditions, symptoms, tests, treatment, medicines, procedures, prognosis and prevention.
Specific questions, not bulk lists. Requests to list or export large sets ("all drugs that…", "every condition with…") are declined. Ask about the situation you are dealing with.
Not a patient-specific recommendation engine. For a particular patient use Decision cards or the care model.

Refusals

A declined question returns 200 with message.content: null and message.refusal set, and is not billed:

JSON

{"role": "assistant", "content": null,
 "refusal": "This model answers clinical questions only (conditions, symptoms, tests, treatment, medicines, procedures). Please ask a clinical question.",
 "augsidius": {"refusal_code": "not_clinical", "citations": []}}

`refusal_code`	Why
`not_clinical`	The question is not about human health or clinical care.
`bulk_request`	The question asks for a list or dump rather than a specific clinical answer.

When streaming, the refusal arrives in delta.refusal on the final chunk.

Limits and price


Input	Up to 2,000 tokens (question plus earlier turns)
Typical time	7–12 seconds; text starts streaming after 3–5 seconds
Price	8 units per call per 1,000 input tokens — see Plans & billing

SDKs & examples

Care models