More API

Enrich

Fill the empty cells of a table from the web. Send rows and the columns you want filled; every cell comes back with a value, a status and the page it was read from.

Best for: lead lists, company tables, catalog gaps
Route: /v1/enrich (job on the hosted API)
Start with: a website column

How it works

For each row, fastCRW:

  1. Reads the row's own website first, when the row has a URL or a domain in any value.
  2. Searches for the row's entity, using its other values and your context, then reads the top results together with their search summaries.
  3. If a column is still missing, searches once more for that column and reads the top new results. A row uses at most two searches.
  4. Merges per cell: a value a page states wins over one seen only in a search summary.

A cell is supported only when a page states the value word for word; the cell names that page and quotes the sentence. A value seen only in a search result's summary comes back unverified, citing that result. Only sources that name the row's entity are read, and the model is told to set aside a namesake; that is a guard, not a guarantee.

Hosted API

On api.fastcrw.com, enrich runs as a job, so a table can be as large as your plan's job limit:

POST https://api.fastcrw.com/v1/enrich
Authorization: Bearer YOUR_API_KEY

The body is the request described below, plus an optional budget (maxCredits, maxDurationMs). The answer is 202:

{ "success": true, "id": "cm3...", "status": "accepted", "expiresAt": "2026-11-08T10:00:00.000Z",
  "budget": { "maxCredits": 222, "maxDurationMs": 600000, "maxSubrequests": 120 } }

Poll GET /v1/enrich/{id}. It returns every row in request order: rows waiting their turn are "status": "pending" and rows being filled right now are "running", both with empty cells; finished rows are the row objects described below plus creditsUsed.

{
  "success": true,
  "id": "cm3...",
  "status": "running",
  "budget": { "maxCredits": 222, "creditsUsed": 18, "maxDurationMs": 600000, "maxSubrequests": 120 },
  "data": { "rows": [ { "index": 0, "rowId": "acme-1", "status": "completed", "cells": { }, "creditsUsed": 18 },
                      { "index": 1, "rowId": "acme-2", "status": "pending", "cells": {} } ] }
}
Job status Meaning
accepted, running Rows are being filled.
completed Every row has a result.
partial The job stopped early (stopReason: budget_exhausted, deadline_exceeded or subrequest_limit). Rows still pending were not run: send them as a new job, matched back by rowId.
cancelled Stopped by DELETE /v1/enrich/{id}. Rows that finished keep their results.

Billing follows the usual rules: each page read costs what a scrape of it costs, each search what a search costs, plus the model usage for that row. A row that fails is not billed. The default budget covers the most the rows could cost; pass a smaller maxCredits to cap a job, and expect partial when it runs out. Enrich is available on paid plans; the hosted API runs it on its managed model, so llmApiKey and related fields are not accepted there.

The hosted MCP server exposes it as crw_enrich for up to 10 rows: it returns the table when it is done, or the rows so far plus the job id to call again with.

Request

POST /v1/enrich
{
  "rows": [
    { "id": "acme-1", "company": "Stripe", "website": "stripe.com", "ceo": null },
    { "id": "acme-2", "company": "Linear", "website": "linear.app", "ceo": "" }
  ],
  "columns": [
    { "name": "ceo", "description": "Current CEO, full name", "type": "text" },
    { "name": "employees", "description": "Approximate employee count", "type": "number" },
    { "name": "stage", "description": "Latest funding stage", "type": "options",
      "options": ["Seed", "Series A", "Series B", "Later", "Public"] }
  ],
  "context": "B2B software companies"
}
Field Notes
rows Flat objects: at most 3 per call on a self-hosted server, up to your plan's job limit on the hosted API. Values that are not target columns describe the row's entity. An id (string or number) is echoed back as rowId.
columns At most 10. name is the row key to fill, description says what belongs in it.
columns[].type text (default), number, url or options. For options, a value that matches none of the labels comes back notFound. Dates are text, as written on the page.
context Optional. One line about the table, used to tell apart entities with the same name.
overwrite Optional, default false. A cell is filled only when its value is missing, null or ""; true refills every listed column.

Response

{
  "success": true,
  "status": "completed",
  "data": {
    "rows": [
      {
        "index": 0,
        "rowId": "acme-1",
        "status": "completed",
        "cells": {
          "ceo": {
            "value": "Patrick Collison",
            "status": "supported",
            "confidence": "high",
            "citations": [
              { "url": "https://stripe.com/about", "excerpt": "Patrick Collison, CEO" }
            ]
          },
          "employees": { "value": null, "status": "notFound", "citations": [] }
        },
        "usage": { "pages": ["https://stripe.com/about"], "searches": 1 }
      }
    ]
  }
}

Rows come back in request order. Keep only what you trust:

const filled = row => Object.fromEntries(
  Object.entries(row.cells)
    .filter(([, cell]) => cell.status === "supported")
    .map(([name, cell]) => [name, cell.value])
);

Cell status

status Meaning
supported The value appears word for word in the quoted excerpt on the cited page.
unverified The cited page was read and gave this value, but no exact sentence could be matched. Check it before you rely on it.
notFound No page that was read stated this value. value is null.

A target column that is missing from cells was never reached, usually because the row ran out of time. That is different from notFound. A row with no blank target cells comes back completed with empty cells.

Cells may carry extra evidence fields (basisVersion, field, sourceHash, sourceTextKind). usage.llmUsage reports the extraction token usage. Both are informational and may change.

Row status

status Meaning
completed Every target column was looked for.
incomplete The row ran out of time (stopReason: "timeout"). Cells it did not reach are absent. Send the row again.
failed No page could be read for the row (error says why; stopReason: "timeout" when time ran out first). Send it again, or add a website value.

Larger tables on a self-hosted server

On the hosted API, send the whole table as one job. A self-hosted server answers synchronously and takes 3 rows per call, so chunk bigger tables and join the results back by rowId:

const results = [];
for (let i = 0; i < rows.length; i += 3) {
  const res = await fetch(`${BASE}/v1/enrich`, {
    method: "POST",
    headers: { "Content-Type": "application/json", Authorization: `Bearer ${KEY}` },
    body: JSON.stringify({ rows: rows.slice(i, i + 3), columns, context }),
  }).then(r => r.json());
  results.push(...res.data.rows);
}

Tips

  • Include a website column. The row's own site is read first, and it is the strongest way to keep a common name from matching the wrong company.
  • Describe each column the way the fact is written on a page ("Current CEO, full name"), not as a database field.
  • Ask for facts pages state. Inferred values ("is this company a good fit?") come back notFound.

Self-hosting

/v1/enrich needs an LLM ([extraction.llm], or llmApiKey per request) and works best with a search backend configured. Without search, only rows that carry their own URL get cells. GET /v1/capabilities reports both under enrich.

The MCP server exposes the same call as the crw_enrich tool.