Run Qdrant recommend, discover, and context queries through the Query API using positive/negative examples

Documented steps

1. All three modes use one endpoint: POST /collections/{collection_name}/points/query.
2. Recommend: {"query":{"recommend":{"positive":[100,231],"negative":[718,[0.2,0.3,0.4,0.5]],"strategy":"average_vector"}},"limit":3}. positive/negative accept point IDs or raw vectors interchangeably.
3. strategy 'average_vector' (default) builds a single search vector as avg_positive + avg_positive - avg_negative; performance is on par with a regular search.
4. strategy 'best_score' (v1.6.0+) scores each candidate against every example with sigmoid normalization, scales linearly with example count, supports negative-only examples, and needs ef raised to 64+ for accuracy.
5. strategy 'sum_scores' aggregates per-example similarity sums across positive and negative sets and also supports negative-only examples.
6. Target a non-default named vector with "using":"image"; pull example vectors from another collection with "lookup_from":{"collection":...,"vector":...}; batch many recommendations via POST /collections/{collection_name}/points/query/batch.
7. Discover (v1.7+): {"query":{"discover":{"target":[0.2,0.1,0.9,0.7],"context":[{"positive":100,"negative":718},{"positive":200,"negative":300}]}},"limit":10}. Each context pair splits the vector space into a positive and a negative zone.
8. Context-only search (no target): {"query":{"context":[{"positive":100,"negative":718}]},"limit":10}. Scoring is loss-based where 0.0 is optimal, and results spread across the permitted zone rather than clustering around one point.
9. Reference: https://qdrant.tech/documentation/search/explore/