Extract structured data with Go

Problem: you need a precise answer about a company (founders, funding, stack, decision makers) without creating a table. Solution: POST /v1/go/run with an output_schema. The Go agent searches the web and reads pages. It can also call every Retriever service: LinkedIn and Sales Navigator people search, LinkedIn posts, company data, emails and phone numbers. It chooses the tools itself; say in instructions which sources you want if it matters. Each service it calls is billed at its price, in addition to the model’s work: usage.credits.breakdown lists them. Services called by a job that fails or is cancelled stay billed, and the error response does not list them. A question that needs a phone number costs much more than a web search.

import os, requests

API = os.environ["RETRIEVER_API_URL"]
HEADERS = {"Authorization": f"Bearer {os.environ['RETRIEVER_API_KEY']}"}

res = requests.post(f"{API}/v1/go/run", headers=HEADERS, timeout=180, json={
    "query": "Qonto's latest funding round: amount, date, lead investors.",
    "instructions": "Cross-check at least two reliable sources. Include source URLs.",
    "model": "fast",
    "output_schema": {
        "type": "object",
        "required": ["amount_eur", "date", "investors", "sources"],
        "properties": {
            "amount_eur": {"type": ["number", "null"]},
            "date": {"type": ["string", "null"]},
            "investors": {"type": "array", "items": {"type": "string"}},
            "sources": {"type": "array", "items": {"type": "string"}},
        },
    },
})
body = res.json()
if not res.ok:
    raise RuntimeError(f"{res.status_code} {body['error']['code']}: {body['error']['message']}")
print(body["value"], "credits:", body["usage"]["credits"]["total"])

Beyond the web: people on LinkedIn The same request works for questions that need Retriever’s services. Here the agent searches LinkedIn for the people and confirms their current role. This search can take a few minutes, so it uses the asynchronous POST /v1/go/runs: it returns a run_id, then poll GET /v1/go/runs/{id}.

curl -X POST "$RETRIEVER_API_URL/v1/go/runs" \
  -H "Authorization: Bearer $RETRIEVER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "Who manages the vehicle fleet at Decathlon France (decathlon.fr)? Current employees only.",
    "instructions": "Search people on LinkedIn at this company, then confirm each current title. At most 5 people.",
    "output_schema": {
      "type": "object",
      "required": ["people"],
      "properties": {
        "people": {
          "type": "array",
          "items": {
            "type": "object",
            "required": ["name", "title", "linkedin_url"],
            "properties": {
              "name": {"type": "string"},
              "title": {"type": "string"},
              "linkedin_url": {"type": "string"}
            }
          }
        }
      }
    }
  }'

Good to know

Beyond about 30 s, prefer POST /v1/go/runs then GET /v1/go/runs/{id}: you avoid HTTP timeouts.
Do not auto-replay POST /v1/go/run after a timeout: the search may have succeeded and been billed.
402 insufficient_credits: check GET /credits before large batches. A search starts only with 1 credit available: that amount is held for the model’s reasoning before any work, and what the search does not use is given back. Below it, POST /v1/go/run answers 402, and a run started with POST /v1/go/runs ends failed with insufficient_credits. Each search holds its amount while it runs, so parallel searches need that much each. Services a search calls (a phone lookup, for example) are billed on top of that amount as they run, so plan for them before a large batch.
For hundreds of companies, import them into a table and use a run (recipe 1).

Get started

Model Context Protocol

Concepts

Recipes

Extract structured data with Go