Paginate Qdrant results with the scroll API and run cheap-then-precise multi-stage rescoring with prefetch and quantization oversampling

Documented steps

1. Scroll all points matching a filter: POST /collections/{collection_name}/points/scroll with {"filter":{"must":[{"key":"color","match":{"value":"red"}}]},"limit":100,"with_payload":true,"with_vector":false}.
2. Results are sorted by ID. Take next_page_offset from the response and pass it as "offset" in the next request; a null next_page_offset means pagination is complete.
3. order_by (v1.8.0+) sorts scroll results by a payload key but disables next_page_offset pagination. To page deep with ordering, combine order_by with a start_from value plus a filter.
4. For multi-stage retrieval, POST /collections/{collection_name}/points/query with a "prefetch" block: when prefetch is present, Qdrant runs it first, then applies the top-level query over only those candidates.
5. Cheap-to-precise: {"prefetch":{"query":[1,23,45,67],"using":"quantized","limit":1000},"query":[0.01,0.45,0.67],"using":"full","limit":10}.
6. ColBERT/multivector rescoring: {"prefetch":{"query":[0.01,0.45,0.67],"limit":100},"query":[[0.1,0.2,0.32],[0.2,0.1,0.52]],"using":"colbert","limit":10}.
7. prefetch blocks nest arbitrarily deep for multi-hop pipelines, e.g. mrl_byte -> full -> colbert.
8. score_threshold is a top-level /points/query field and is not honored inside prefetch sub-queries.
9. Tune quantized rescoring: {"params":{"quantization":{"rescore":true,"oversampling":2.0}}}. Oversampling multiplies how many extra candidates are pre-selected via the quantized index before final rescoring (oversampling 2.4 with limit 100 pre-selects 240).
10. Reference: https://qdrant.tech/documentation/manage-data/points/
11. Reference: https://qdrant.tech/documentation/search/hybrid-queries/
12. Reference: https://qdrant.tech/documentation/manage-data/quantization/