@recourse-ai/store-d1

The recourse Store, backed by Cloudflare D1.

import { createChatHandler } from '@recourse-ai/core/server'
import { d1Store } from '@recourse-ai/store-d1'

export default {
  async fetch(request: Request, env: Env) {
    return createChatHandler({ index, store: d1Store({ db: env.DB }) })(request)
  },
}

// wrangler.jsonc
{ "d1_databases": [{ "binding": "DB", "database_name": "recourse", "database_id": "..." }] }

Choosing this over Postgres

The deployment, not the database. A Worker reaches D1 through a binding: no connection pool, no credential, nothing to exhaust. On a runtime where every instance opening a pool is the failure mode, not having pools is the feature.

The bill

| | Free | Paid | | --- | --- | --- | | Database size | 500 MB | 10 GB | | Queries per Worker invocation | 50 | 1,000 | | Databases per account | 10 | 50,000 | | Point-in-time restore | 7 days | 30 days |

The one that will catch you out is 50 queries per invocation on the free tier, because it is per request rather than per day. getConversation is two; a turn that also writes messages and lists tickets adds more. Count them if you are near it.

Cloudflare's own guidance is many small databases rather than one large one, which suits this product: one per business, and the size ceiling stops being a ceiling.

Differences from the Postgres store

Same interface, same behaviour suite, different SQL underneath.

• INTEGER PRIMARY KEY AUTOINCREMENT rather than a sequence. The ticket-number race is still fixed, SQLite assigns inside the insert rather than the application reading the highest and adding one.
• FTS5 for ticket search. Its query syntax treats punctuation as operators, so every word is quoted before searching; otherwise an apostrophe in a search box is a syntax error rather than a search.
• JSON in TEXT columns. SQLite has JSON functions but no JSON type, and everything here is read whole rather than queried into.
• No advisory lock around migration. A D1 database is a single Durable Object and runs one statement at a time, so the race the Postgres version guards against cannot happen.

Running the tests

pnpm --filter @recourse-ai/store-d1 test

No database and no dependency: the tests drive the store through a shim over node:sqlite, which Node 22 ships built in. D1 is SQLite, so this exercises the same SQL the Worker will run.

What that leaves unproven, honestly: the binding itself, the network, and the per-invocation query limit. Those need wrangler dev --local with a real D1 binding, and a deploy needs a Cloudflare account.