Qdrant: Picking a Vector Database Without the Hype

Choosing a vector database has become a benchmark-quoting contest, and the benchmarks measure the thing that matters least.

Raw nearest-neighbour speed is a solved problem. Every serious option is fast enough. What separates them in a real application is what happens when you need to search and filter.

Filtered search is the actual requirement

Nobody searches an entire corpus. Real queries look like: find documents similar to this, but only from this client, only from this year, only published.

Naive implementations retrieve the top hundred by similarity and then filter, which means if none of the top hundred match your filter you get nothing back. The database has to apply constraints during the search, not after.

Qdrant handles this properly, with indexed payload filters that participate in the search itself. That is the feature I would pick on, and it is the one benchmarks rarely test.

client.search(
    collection_name="docs",
    query_vector=embedding,
    query_filter=Filter(must=[
        FieldCondition(key="client_id", match=MatchValue(value="tulip")),
        FieldCondition(key="published", match=MatchValue(value=True))
    ]),
    limit=5
)

The other things that actually matter

Operational simplicity. A single binary or container you can run on the same box as the app is worth a great deal to a small team. A cluster you must operate is a cost you pay every month.

Payload storage. Keeping the source text next to the vector avoids a second lookup in another store for every result.

Snapshots. Re-embedding a large corpus is slow and expensive. Backups are not optional.

Sensible defaults. You should not need to tune index parameters to get good recall on day one.

When you do not need one at all

If you have a few thousand chunks, a vector index in Postgres via pgvector is enough, and it saves you an entire service to deploy, monitor and back up. I have shipped client features on exactly that.

Reach for a dedicated vector database when the collection is large, when filtering is complex, or when the workload genuinely justifies its own service. "It seemed more serious" is not a technical reason.

Resources

• Repo: qdrant/qdrant
• Docs: qdrant.tech/documentation
• Video walkthroughs: YouTube: Qdrant tutorial
• Feeding it: LlamaIndex chunking