EvaEsi Reasoning | EvaEsi Dev

Overview

Reasoning is the deeper pipeline. It gathers evidence first, then composes the final answer from the collected sources. Use it when the question needs chained inference, comparison across sources, or a defensible position rather than a quick reply.

Reasoning supports the same streaming and citation surface as Instant. Output sizes can be twice as large to accommodate longer arguments.

Endpoint

POST https://evaesi.dev/api/v1/responses

Pass "model": "reasoning" in the JSON body to route through this surface.

When To Use Reasoning vs Instant

Question shape	Pick
Casual chat, definitions, short factual lookups	`instant`
News summary with citations	`instant` with `web_search: true`
Multi-step analysis or argumentation	`reasoning`
Synthesis across multiple sources	`reasoning`
"Why" questions requiring chained inference	`reasoning`

Limits

Field	Value
Max input rungs	`100,000`
Max output rungs	`64,000`
Context window	`4,096` model bytes (billing rungs are separate: 4,096 input bytes meter as 5,734 rungs here)
Streaming	Yes
Batch	No. The route is mounted but returns `503 EVA_UNAVAILABLE`
Tools	No

Billing

Reasoning is billed at 1.4× the Instant rate on both input and output rungs. The multiplier is applied server-side before any usage row is written, so the rungs reported on usage and on the x-evaesi-*-rungs headers already include the multiplier.

Example: a 100-byte prompt that produces 500 bytes of output charges 140 input rungs and 700 output rungs. A maximal 100,000-byte prompt meters as 140,000 input rungs.

The request

Same JSON shape as Instant. The web_search flag is accepted but the evidence-gather stage runs whenever the underlying pipeline determines it is needed.

Field	Type	Required	Description
`prompt`	string	Yes	Prompt text. Max 100,000 bytes.
`model`	string	Yes	Pass `"reasoning"` to route here.
`max_output_rungs`	u64?	No	Output cap, validated against the model's 64,000 limit (`400 INVALID_REQUEST` above it). Falls back to 64,000.
`stream`	bool?	No	Server-sent event streaming.

Response

{
  "request_id": "req_e201...",
  "content": "...",
  "sources": [
    {
      "url": "https://example.org/report",
      "domain": "example.org",
      "title": "Example report",
      "tier": "t1_authoritative",
      "retrieved_at": "2026-05-12T09:00:00Z",
      "snippet": "Quoted evidence...",
      "stance": "supports"
    }
  ],
  "sources_url": "/api/v1/sources/req_e201...",
  "evidence_verdict": "supported",
  "usage": {
    "input_rungs": 168,
    "output_rungs": 980,
    "total_rungs": 1148
  },
  "model": "reasoning"
}

Notice 168 input rungs on a 120-byte prompt, that is the 1.4× multiplier showing up on the reported usage.

Examples

cURL

curl -X POST https://evaesi.dev/api/v1/responses \
    -H "Authorization: Bearer esi_dev_YOUR_KEY" \
    -H "Content-Type: application/json" \
  -d '{
    "prompt": "Compare the 2024 and 2026 UK energy policy reforms.",
    "model": "reasoning",
    "max_output_rungs": 2048
  }'

Python

import requests

response = requests.post(
    "https://evaesi.dev/api/v1/responses",
    headers={"Authorization": "Bearer esi_dev_YOUR_KEY"},
    json={
        "prompt": "Compare the 2024 and 2026 UK energy policy reforms.",
        "model": "reasoning",
        "max_output_rungs": 2048,
    },
)

data = response.json()
print(data["content"])
print(f"sources cited: {len(data['sources'])}")
print(f"verdict: {data.get('evidence_verdict')}")

JavaScript

const response = await fetch("https://evaesi.dev/api/v1/responses", {
  method: "POST",
  headers: {
    "Authorization": "Bearer esi_dev_YOUR_KEY",
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    prompt: "Compare the 2024 and 2026 UK energy policy reforms.",
    model: "reasoning",
    max_output_rungs: 2048,
  }),
});

const data = await response.json();
console.log(data.content);
console.log(`sources: ${data.sources.length}, verdict: ${data.evidence_verdict}`);

Start from a live reference page.