civic-data-adapters

TypeScript adapters for municipal and civic government data, extracted from a production local-news pipeline. One registry, normalized records, bring-your-own storage.

npm install civic-data-adapters

What it covers

Eight working providers:

Provider	Data
legistar	Council/board meetings and agenda items (Legistar API)
socrata	Any Socrata-hosted dataset: blotters, permits, violations, contracts
blotter_html	Police blotters published as HTML tables (column mapping configurable)
blotter_pdf	Police blotters published as PDFs
blotter_crimewatch	CrimeWatch platform blotters
html_minutes	Free-form HTML meeting minutes (requires an LLM callback, see below)
usaspending	Federal contracts and grants by locality (USASpending API)
nonprofit_explorer	Nonprofit filings by locality (ProPublica API)

Three providers are recognized but stubbed, because they have no public API and need per-site manual setup: granicus, civicplus, boarddocs. The registry logs a warning and returns zeros for them. Contributions welcome; we won't pretend they work.

Design

Adapters know nothing about your database or your app. Each one takes a locality, a source config, and a context, then hands normalized CivicRecords to your sink:

import { runAdapter, MemorySink, type CivicAdapterMeta } from "civic-data-adapters";

const sink = new MemorySink(); // or implement CivicRecordSink over your DB

const result = await runAdapter(
  { name: "Springfield", state: "IL" },
  { id: "src-1", url: "https://data.example.gov" },
  {
    provider: "socrata",
    url: "https://data.example.gov",
    resourceId: "abcd-1234",
    recordType: "police_blotter",
    dateField: "incident_date",
  } satisfies CivicAdapterMeta,
  { sink },
);
// result: { inserted, skipped }

Every record carries a stable dedupeKey (provider plus the source's natural id plus date), so sinks can upsert idempotently. MemorySink ships for tests and dry runs; a real deployment implements the one-method CivicRecordSink interface over its own storage.

Other things the context controls: fetch override (tests, proxies), userAgent, and robots.txt checking, which is on by default and fails open. Dates pass through sanitizeCivicDate, which rejects unparseable values and anything more than a year in the future, because open-data date fields do contain typos and a bad future date can make a "recent activity" view permanently wrong. That one comes from production experience.

The LLM callback

Two adapters touch unstructured text. legistar can summarize agenda items, and html_minutes cannot work at all without a parser for free-form minutes. Both use one optional hook:

ctx.summarize = (systemPrompt, text) => myLlm(systemPrompt, text); // returns Promise<string>

Without it, legistar falls back to title/action text and html_minutes skips with a warning. Bring any LLM. If you want governance around that spend (caps, usage ledger, failover), llm-governance-gateway's runText wires in directly, but nothing here requires it:

ctx.summarize = async (system, text) =>
  (await gw.runText({ slug: "minutes", promptBody: text, system, cache: false, ... })).text;

Discovery (v0.2)

Point at a city and find its civic sources. Discovery proposes candidates through an LLM callback, then verifies every one by running the real adapter parsers against it. A returned source has actually yielded data, right now, not "the model thinks this looks right":

import { discover } from "civic-data-adapters";

const sources = await discover(
  { name: "Springfield", state: "IL" },
  {
    generate: myWebSearchGroundedLlm, // (systemPrompt, prompt) => Promise<string>
    socrataPortalUrl: "https://data.springfield.il.gov", // optional
  },
);
// Each result: { provider, name, meta, evidence }
// evidence reads like "parsed 12 HTML blotter rows just now"
// meta drops straight into runAdapter — after a human approves it.

Individual entry points also exist: discoverBlotter, discoverMeetingPortal, and searchSocrataCatalog (the catalog one needs no LLM at all).

Two things stated plainly. First, the prompts ask for real, current URLs, so generate should be backed by web-search-grounded generation; an ungrounded model will hallucinate candidates, which verification then rejects — safe, but you'll get nothing. Second, discovery proposes and humans approve; nothing here auto-provisions scraper targets.

Roadmap

See ROADMAP.md: recorded-fixture tests for the blotter parsers, politeness controls beyond robots.txt, and real Granicus/CivicPlus/BoardDocs adapters when someone does the per-platform work.

License

Apache-2.0