November 15, 2026
Finding the Right Podcast Episode: How Search Works on HittaPodd.se
How we use Meilisearch on HittaPodd.se so listeners find podcasts and episodes quickly, even with typos, and how we want to use it for recommendations next.
Search is the front door of a podcast app. Someone has a show or a topic in mind, they type a few letters, and they expect the right thing to show up. If it doesn't, they don't think "the search index is stale". They think "this app doesn't have it".
On HittaPodd.se we use Meilisearch for this. It sits behind our API and is used by the web app. This post is about why we added it, and what it does for listeners.
What was wrong before
Search used to be a database query that looked for your text inside titles. It worked, but badly in three ways:
- It was slow where it mattered. Episodes are the biggest part of the catalogue. On our measurements with about 345,000 episodes, the database search took around 1.7 seconds. The search engine answers in about 10 milliseconds. A search box that waits almost two seconds after every keystroke feels broken.
- It didn't forgive typos. Type
fotbolsvikenwith oneland you got nothing, even though the show exists. - The order was arbitrary. A database
LIKEquery finds matches but doesn't say which is the best one.
What listeners get now
Fast enough to search as you type. We lowered the delay before a search starts from 400 to 150 milliseconds. That only makes sense because answers come back quickly. Results appear while you're still typing, in a popup you can open from anywhere with Cmd/Ctrl+K.
Typos are forgiven. A title that's one letter off is still found. We tuned this on purpose, though. By default the engine forgives a typo from 5 letters, and at that length krimm started matching krim. We changed it to forgive one typo from 6 letters and two from 12, so short words stay precise.
Better order. Of equally good matches, a podcast with more episodes comes first, and a newer episode comes before an older one. We also drop weak matches. Without that, searching for kolbäck would list every episode that happens to mention the village of Kolback somewhere deep in its description.
One search across everything. The same box finds podcasts, episodes, collections and focus areas. There's also a results page at /hitta where podcasts show as cards and episodes as rows with cover, show name, snippet and date. The address contains your search, so you can share it.
Newly published episodes show up quickly. A new episode of a show you follow shouldn't be missing from search for a day. A small job adds new episodes to the index every ten minutes. A full rebuild runs nightly to catch everything else.
Where it still falls short
We try to be honest about trade-offs in the docs, and the same applies here.
It matches the start of words, not the middle. fotboll finds Fotbollspodden. podden does not. The old database query would have found it. That's the price of the typo tolerance, and it hurts with Swedish compound words. I think the trade is right, since typos are more common than searching for the end of a word. But some listeners will notice.
New things take a little while. Not everything is picked up by the ten-minute job. If a feed changes an episode's title, that waits for the nightly rebuild.
It uses real resources. The index for about 260,000 episodes takes around 2 GB of disk, and a rebuild needs room for the old and new index side by side. That's fine for us, but it's not free.
Why Meilisearch
We looked at two open source engines: Typesense and Meilisearch. Both are fast, both forgive typos, and both can run on your own server. Either would probably have worked for us. We went with Meilisearch.
What we ended up relying on, and what makes it a good fit for HittaPodd:
- Typo tolerance we can tune. We set how long a word must be before a typo is forgiven, which fixed short words matching too loosely.
- A ranking score for every hit. That let us drop weak matches, like a near miss buried in a long description.
- Language settings. We can tell the engine which language a field is written in, which matters for Swedish.
- A plain HTTP interface. Our client is small and has no extra dependency. That keeps the search code easy to read and replace.
I'd say the same to anyone choosing between the two: try both on your own data. Search quality depends on your content, and a day of testing tells you more than a comparison table.
And compared with Algolia
The hosted name most people know is Algolia. It's good and easy to start with, and for many teams it's the right call. If you're willing to self-host, though, Meilisearch or Typesense is a good alternative:
- It's ours. The data stays on our own server. There's no third party in the path of every search a listener makes.
- The cost is predictable. We pay for a server we already run, not per search or per record. With a catalogue of hundreds of thousands of episodes, that matters.
- It's small to run. For us it's one container with a memory cap and no public port. Only our own services can reach it.
The honest other side is that you take on the work. You size the server, keep the index up to date, and handle upgrades yourself. A hosted service does all of that for you, and for a small team that may be worth the price. Meilisearch also offers a hosted version if you want the engine without running it.
Our setup makes self-hosting lower-risk than it sounds, because the index is only a fast copy. If it's lost, we rebuild it from the database.
The safety net
The part I like most is what happens when the search engine is down. The answer is: nothing visible. The API quietly falls back to the old database search, and the listener gets slower, less forgiving results instead of an error. We only see it in the logs.
Two design choices make that safe.
The engine only decides what matches and in what order. The actual podcast and episode data is always read from our database. So if an admin hides an episode, or a podcast is removed, or a collection is made private, it disappears from search results right away. The index doesn't have to be rebuilt first. The result looks the same whether it came from the engine or from the fallback.
The index can always be rebuilt. Meilisearch is a fast copy of what's in the database, never the source of truth. The nightly rebuild builds a fresh index on the side and swaps it in, so a search never hits one that's half finished. Upgrading Meilisearch means wiping the index and rebuilding it. That sounds drastic and is actually boring, which is what we want.
What's next: finding what you didn't know to search for
Search only helps people who already know what they're looking for. Some of the best listening comes from the show you'd never have typed in. That's the next job for the same engine: recommendations and personalization.
None of this is built yet, so this is the direction and not a feature list. But Meilisearch is a good fit for it, and here is what we want it to make possible for listeners:
- "If you liked this, try that." An episode or podcast page that suggests things close to it in topic, not just things with a similar name. We already find similar podcasts today. A search engine that understands words and ranking lets us do it for episodes too, and do it fast enough to show on every page.
- A start page that knows you. What you've listened to, saved and followed can shape what you see first, so you spend less time scrolling through things that don't interest you.
- Search that leans your way. Of two equally good matches, the one closer to what you usually listen to can come first. Someone who follows football podcasts probably means the sports show when they type a short, ambiguous word.
- Discovery beyond the obvious. Recommendations that only repeat the most popular shows aren't much help. We'd like to surface good, smaller podcasts that match your taste and that you'd otherwise miss.
Personalization comes with a responsibility. We want it to be something you can see and control, not a black box. That means being clear about what's used, keeping it to what a listener has chosen to do on HittaPodd, and always leaving a plain, non-personalized way to browse and search. A recommendation that feels creepy is worse than none.
We'll write about it as it takes shape, including what doesn't work. In the meantime, I'd also like to look at which searches return nothing, since those show what listeners want and we don't have. Compound words are the other thing on my mind, since they're where the engine is weakest for Swedish.
If a search on HittaPodd doesn't find what you expected, tell us. It's the quickest way for us to find the next thing to fix.