I'm unhappy with Qdrant memory usage

Got it — Qdrant’s RAM usage can be surprising, especially at larger scales or with many payload fields / vectors.

If you want, I can help you reduce it. Common memory drivers in Qdrant are:

• Vector index (HNSW) in RAM
• Payload indexes for filtered search
• Multiple vectors per point
• Large payloads stored alongside points
• High replication / many shards
• Too much segment fragmentation from frequent upserts/deletes

Typical ways to cut memory:

1. Use quantization

   • Scalar or product quantization can reduce vector memory substantially.

2. Tune HNSW

   • Lower m
   • Lower ef_construct
   • Sometimes reduce search ef if recall requirements allow

3. Reduce payload indexing

   • Only index fields you actually filter on
   • Avoid indexing high-cardinality fields unless needed

4. Use disk-backed / on-disk options

   • Keep vectors or payloads on disk where possible, depending on your version/config.

5. Compact and optimize collections

   • Merge segments, remove tombstones, and reduce fragmentation.

6. Shrink vector size or count

   • Fewer dimensions
   • One vector instead of multiple where possible
   • Store derived metadata outside Qdrant

7. Review replication and shard settings

   • They multiply memory footprint.

If you want, send me:

• your collection config,
• vector size and count,
• number of points,
• payload fields you filter on,
• and whether you use HNSW / quantization / on-disk settings,

and I can suggest the most likely memory wins.