Mercury-Decide
Mercury-DecideFast structured decisions from Inception. Pass JSON state and a set of questions; Mercury-Decide returns choices, scores, and checks in one call.
Model details
Mercury-Decide is Inception's purpose-built decision model. Instead of generating free-form text, it answers structured questions about your data and returns machine-readable decisions, so your application gets a choice, a score or a check it can act on directly, with no parsing of chat output.
You send a JSON state (the record, event or context to evaluate) and a set of named questions. Mercury-Decide answers each one in a single call:
Choice: pick one option from a defined set of criteria (for example, route a ticket to billing, technical or sales)
Score: select a level on an ordered scale you define (for example, a 0–2 rating)
Checks: yes/no-style questions about the state (for example, whether a field is present or null)
Key capabilities:
Multiple decisions per request: ask several questions about the same state in one call
Probabilities, not just answers: returns per-option log-probabilities, so you can set confidence thresholds or route low-confidence cases for review
64K-token context for long records, documents or conversation histories
Mercury-Decide uses a dedicated Decisions API (/v1/decisions) rather than the chat completions format. It's best suited to high-volume, latency-sensitive classification and judgment tasks where answers need to be structured and auditable, such as content and policy moderation, ticket triage and routing, data validation and extraction QA, eligibility and compliance checks, and LLM-as-judge evaluation.
Here's an example calling it via the API:
1import os
2import requests
3
4response = requests.post(
5 "https://inference.baseten.co/v1/decisions",
6 headers={"Authorization": f"Api-Key {os.environ['BASETEN_API_KEY']}"},
7 json={
8 "model": "inception/mercury-decide",
9 "state": {
10 "ticket": "I was charged twice for my subscription this month.",
11 "customer_tier": "pro",
12 },
13 "questions": {
14 "category": {
15 "type": "choice",
16 "instructions": "Which team should handle this ticket?",
17 "criteria": {"billing": None, "technical": None, "sales": None},
18 },
19 "urgency": {
20 "type": "score",
21 "instructions": "How urgent is this ticket?",
22 "criteria": ["Low", "Medium", "High"],
23 },
24 "is_refund_request": {
25 "type": "noul",
26 "instructions": "Is the customer asking for money back?",
27 },
28 },
29 },
30)
31answers = response.json()["answers"]
32print(answers["category"]["choice"]) # "billing"
33print(answers["urgency"]["score"]) # ~1.8 on a 0–2 scale
34print(answers["is_refund_request"]["noul"]) # ~0.78 probability1{
2 "category": {
3 "type": "choice",
4 "choice": "billing",
5 "probabilities": {
6 "billing": 0.9998,
7 "technical": 0.0001,
8 "sales": 0.0001
9 },
10 "confidence": 0.9997
11 },
12 "urgency": {
13 "type": "score",
14 "score": 1.8,
15 "legend": {
16 "0": "Low",
17 "1": "Medium",
18 "2": "High"
19 },
20 "probabilities": {
21 "0": 0.02,
22 "1": 0.161,
23 "2": 0.818
24 },
25 "confidence": 0.697
26 },
27 "is_refund_request": {
28 "type": "noul",
29 "noul": 0.777
30 }
31}
