Decision and extraction models for runtime

Sage2 extracts structured data and returns decisions with calibrated confidence. Discover the full capabilities of Sage2 below.

Start free Playground Read the docs

DecideRequest: POST /decide with the yes or no question “Does this message ask for a refund?” about a message that asks to refund a duplicate charge. Response: yes, probability of yes 0.999, 104.1 ms on the server.
StreamRequest: WS /v1/stream with the question “Does this need escalation to the on-call engineer now?”, then three appends to one session. Append 1: “Hi, one user on our team can't sign in.” Decision: no, probability of yes 0.009, 110.8 ms on the server. Append 2: “Now nobody in our EU workspace can sign in.” Decision: yes, probability of yes 0.991, 76.9 ms on the server. Append 3: “Checkout is failing for every customer and we are losing orders.” Decision: yes, probability of yes 0.999, 131.2 ms on the server.
ExtractRequest: POST /v1/extract with the text “I can pay $450/mo” and the money field monthly_limit. Response: 450 USD, 182.1 ms on the server.

POST/decide

{ "question": { "kind": "yesno", "id": "refund",    "instructions": "Does this message ask for a refund?" },  "content": "Hi, I was charged twice for order #1048…" }{ "question": {    "kind": "yesno", "id": "refund",    "instructions":      "Does this message ask for a refund?" },  "content": "Hi, I was charged twice…" }

{ "id": "refund", "kind": "yesno",  "result": { "answer": "yes", "probability": 0.999 },  "meta": { "latency_ms": 104.1 } }{ "id": "refund", "kind": "yesno",  "result": { "answer": "yes",    "probability": 0.999 },  "meta": { "latency_ms": 104.1 } }

WS/v1/stream

→ { "type": "append",    "text": "Hi, one user on our team can't sign in." }→ { "type": "append", "text":    "Hi, one user on our team can't sign in." }→ { "type": "append",    "text": " Now nobody in our EU workspace can sign in." }→ { "type": "append", "text":    " Now nobody in our EU workspace can sign in." }→ { "type": "append",    "text": " Checkout is failing for every customer and we are losing orders." }→ { "type": "append", "text":    " Checkout is failing for every customer and we are losing orders." }

← { "type": "decision", "latency_ms": 110.8,    "answers": { "escalate": { "result":      { "answer": "no", "probability": 0.009 } } } }← { "type": "decision",    "latency_ms": 110.8,    "answers": { "escalate": {      "result": { "answer": "no",        "probability": 0.009 } } } }← { "type": "decision", "latency_ms": 76.9,    "answers": { "escalate": { "result":      { "answer": "yes", "probability": 0.991 } } } }← { "type": "decision",    "latency_ms": 76.9,    "answers": { "escalate": {      "result": { "answer": "yes",        "probability": 0.991 } } } }← { "type": "decision", "latency_ms": 131.2,    "answers": { "escalate": { "result":      { "answer": "yes", "probability": 0.999 } } } }← { "type": "decision",    "latency_ms": 131.2,    "answers": { "escalate": {      "result": { "answer": "yes",        "probability": 0.999 } } } }

POST/v1/extract

{ "content": "I can pay $450/mo",  "extract": { "monthly_limit": { "type": "money",      "definition": "the monthly limit" } } }  "extract": { "monthly_limit": {      "type": "money",      "definition": "the monthly limit" } } }

{ "values": { "monthly_limit": { "amount": 450, "currency": "USD" } }, "meta": { "latency_ms": 182.1 } }

Multimodal

Sage can analyze images.

150ms

Extraction

Server time on short documents.

3×

Faster decisions on progressive context

Streaming is as low as 50 ms per decision.

Choose the decision type to put intelligence at runtime.

Define the question and criteria in natural language. Sage returns a structured result your code can use to select a tool, assign a category, evaluate a condition, or prioritize work.

POST /decide

content

Hi, I was charged twice for order #1048 this morning. Please refund the duplicate charge.

answer

yes

P(yes)

0.999

content · a list of items with ids

A typo on the pricing page.
Search results load in 4 seconds instead of 1.
Checkout returns 500 errors for all customers.
One user cannot change their avatar.

1
Checkout returns 500 errors for all customers.checkout
2
Search results load in 4 seconds instead of 1.search
3
A typo on the pricing page.typo
4
One user cannot change their avatar.avatar

confidence 0.60 · the API returns the ids in order

Explore decision types

Streaming is 3× faster and uses fewer tokens.

Open a streaming session, define your questions, and append text as it arrives. Sage retains the conversation or document server-side and reuses cached prefix computation, so your client sends new text without uploading the full history on every update.

WS /v1/stream

question · sent once when the session starts

Does this need escalation to the on-call engineer now?

current answer

—

P(yes) …

Customer

Hi, one user on our team can't sign in.

no · 0.035Agent

Sorry about that. Can they reset their password from the sign-in page?

no · 0.003Customer

We tried. Now nobody in our EU workspace can sign in.

yes · 0.981Agent

Thanks. I'm checking the EU region status now.

null · 0.609Customer

Checkout is failing for every customer too. We are losing orders.

yes · 0.994Agent

Found it: an expired certificate on the EU gateway. It is renewed now.

no · 0.220Customer

Sign-in and checkout work again. Thanks for the quick fix!

no · 0.041

Decisions in tens of milliseconds
Emit every update, or only changed answers
16,384 tokens of text and questions per session

Read the streaming API

Extract the values your runtime needs.

Define fields for the information your application needs. Sage turns natural language into typed values, including text, amounts, dates, and durations. Extract one value or every matching mention from a document.

POST /v1/extract

content · write your own

typed values

vehicletextmidsize SUVsource “midsize SUV” start 9 · end 20 · p 0.994

monthly_limitmoney450 USDsource “$450” start 31 · end 35 · p 0.966

termdurationnullnot in the text

down_paymentmoneynullnot in the text

Read the docs

Watch our launch video.

A customer wants a midsize SUV for up to $450 a month. Sage extracts those requirements into fields the application uses to find offers.

Build with your existing stack.

Use the Sage API directly or give the API contract to your coding agent. Existing Sage keys work with extraction and streaming. Your application controls tool execution, validation, and fallbacks.

Copy this prompt into your coding agent

Read https://docs.levanto.ai/extract.md and https://docs.levanto.ai/stream.md. Use https://docs.levanto.ai/llms.txt for the other Sage docs. Inspect this repository and identify one workflow where Sage could handle classification, extraction, or decisions over growing context. Explain the proposed change, then implement the selected integration using SAGE_API_KEY from the environment. Handle missing values, uncertain results, and API errors explicitly. Add an evaluation on representative examples before enabling automated actions.

Existing Sage keys work with extraction and streaming.

Get an API key Read the Sage2 docs