Implementing Advanced Query Patterns

# Run N queries and merge results
queries = [embed(t) for t in ["A", "B", "C"]]
all_hits = {}
for qv in queries:
    for m in index.query(vector=qv, top_k=10)["matches"]:
        prev = all_hits.get(m["id"], 0)
        all_hits[m["id"]] = max(prev, m["score"])
top = sorted(all_hits.items(), key=lambda x: -x[1])[:10]

import math, time
now = time.time()
for m in res["matches"]:
    age_days = (now - m["metadata"]["ts"]) / 86400
    m["boosted"] = m["score"] + 0.1 * math.exp(-age_days / 30)
res["matches"].sort(key=lambda m: -m["boosted"])

filter={"doc_id": {"$nin": already_seen_ids}}

filter={"$and": [
    {"in_stock": True},
    {"$or": [{"featured": True}, {"sales": {"$gte": 100}}]},
    {"category": {"$in": ["tech", "books"]}},
]}

filter={"$and": [
    {"lat": {"$gte": 37.7, "$lte": 37.9}},
    {"lng": {"$gte": -122.5, "$lte": -122.3}},
]}

cutoff = int(time.time()) - 30 * 86400
filter={"timestamp": {"$gte": cutoff}}

# Pinecone has no offset; paginate with score cursor
def page(qv, max_score=None, size=20):
    f = {"score": {"$lt": max_score}} if max_score else {}
    res = index.query(vector=qv, top_k=size, filter=f or None,
                      include_metadata=True)
    return res["matches"], res["matches"][-1]["score"] if res["matches"] else None

def mmr(query_vec, candidates, k=5, lam=0.5):
    selected, remaining = [], candidates[:]
    while len(selected) < k and remaining:
        best, best_score = None, -1
        for c in remaining:
            sim_q = cos(query_vec, c["values"])
            sim_sel = max((cos(c["values"], s["values"])
                           for s in selected), default=0)
            score = lam * sim_q - (1 - lam) * sim_sel
            if score > best_score: best, best_score = c, score
        selected.append(best); remaining.remove(best)
    return selected