Scaling a Search Index from 25 to 32,000 Documents in One Evening

The Symptom

Search results for broad queries were suspiciously thin. climate change — a topic with millions of pages — returned a dozen hits, all mediocre. The engine itself answered in milliseconds, health checks were green, and nothing in the logs looked wrong.

The index was simply small. Our primary halaman_web index contained 25 documents, most of them noise: duplicate example domains, a couple of registry pages, crawl test artefacts. Meanwhile a secondary index named uji_coba — created during early dataset experiments and never cleaned up — held 32,272 documents, overwhelmingly English news and feature articles.

Two indexes, one of them forgotten, split the corpus. Queries hit the wrong half.

Why Split Indices Happen

Index sprawl has familiar causes:

• Experiments start with a new index "so we do not break production" — and then the experiment becomes production by accident.

• Schema migrations create index_v2 and the old one never dies.

• Import scripts write to a hardcoded name while the app reads a different one.

The social fix is boring: name indices for their role, document every import target, and delete or archive experiment indices the day the experiment ends.

The Merge

Because both indices shared the same Meilisearch schema (identical field names and types), the migration was a copy, not a transform:

``js
const from = client.index('uji_coba');
const to = client.index('halaman_web');
let offset = 0;

for (;;) {
const res = await from.getDocuments({ offset, limit: 1000 });
const docs = res.results || [];
if (!docs.length) break;
const task = await to.addDocuments(docs);
await client.tasks.waitForTask(task.taskUid);
offset += docs.length;
}
`

• The client returns { results, offset, limit, total } from getDocuments — not a bare array. Treating it as an array yields zero rows and a very confusing "0 documents copied."

• Writes are asynchronous. addDocuments returns a task; waiting on waitForTask (not the removed waitForTask on the client root) prevents the next batch from racing the last.

After the run: halaman_web` held 32,297 documents, and the same broad queries returned full pages of relevant results immediately — no re-ranking, no re-crawling.

What We Changed Afterwards

• A single documented pipeline: every importer writes to the canonical index.

• A startup check comparing configured index names against expected ones.

• Stale experiment indices exported and removed.

• Result-count sanity in monitoring: a "suspiciously few hits" alert for the most common queries.

The Takeaway

Search quality problems are often data-plumbing problems wearing a ranking costume. Before tuning analyzers or embeddings, check how many documents the query is actually searching — a one-line stats call that would have saved us the whole evening of suspicion.