Vector Search — Animate | AI Wisdom

← Back to topic
🎞 ANIMATE · Mode 3 of 5 · +100 XP— HNSW Search
🏗️Create Index
📥Upsert
🔍Query
🎯Filter
📤Results
🎯 Next: Test →
Step 1 / 5
Speed
Build Index
Create HNSW index with M=16, ef_construct=200💡 Analogy
Like building a road network — M is how many roads each city has; ef_construct is how thorough the planners were.
M is graph fan-out per node. ef_construct controls build-time accuracy. Higher = better recall, more RAM.
Console Output
> Collection "docs" created (768d, cosine, HNSW M=16)
●search.py
Build IndexL6–11
1import qdrant_client
2from qdrant_client.models import Distance, VectorParams, PointStruct, Filter, FieldCondition, MatchValue
3 
4client = qdrant_client.QdrantClient(url="http://localhost:6333")
5 
6# 1. Create collection with HNSW index
7client.recreate_collection(
8    collection_name="docs",
9    vectors_config=VectorParams(size=768, distance=Distance.COSINE),
10    hnsw_config={"m": 16, "ef_construct": 200},
11)
12 
13# 2. Upsert vectors with payload
14client.upsert(
15    collection_name="docs",
16    points=[
17        PointStruct(id=i, vector=v.tolist(), payload={"tenant": t, "text": txt})
18        for i, (v, t, txt) in enumerate(rows)
19    ],
20)
21 
22# 3. Query with metadata pre-filter
23results = client.search(
24    collection_name="docs",
25    query_vector=query_vec.tolist(),
26    query_filter=Filter(must=[FieldCondition(key="tenant", match=MatchValue(value=42))]),
27    limit=10,
28    search_params={"hnsw_ef": 64},
29)
30 
31# 4. Surface results
32for r in results:
33    print(f"{r.score:.3f} {r.payload['text'][:60]}")
33 lines
Timeline — 5 Steps
Build Index
Create HNSW index with M=16, ef_construct=200💡 Analogy
Like building a road network — M is how many roads each city has; ef_construct is how thorough the planners were.
M is graph fan-out per node. ef_construct controls build-time accuracy. Higher = better recall, more RAM.
Console Output
> Collection "docs" created (768d, cosine, HNSW M=16)

Was this page helpful?

Discussion

Create HNSW index with M=16, ef_construct=200

💡 Analogy

Like building a road network — M is how many roads each city has; ef_construct is how thorough the planners were.

M is graph fan-out per node. ef_construct controls build-time accuracy. Higher = better recall, more RAM.

Console Output

> Collection "docs" created (768d, cosine, HNSW M=16)

Create HNSW index with M=16, ef_construct=200

💡 Analogy

Like building a road network — M is how many roads each city has; ef_construct is how thorough the planners were.

M is graph fan-out per node. ef_construct controls build-time accuracy. Higher = better recall, more RAM.

Console Output

> Collection "docs" created (768d, cosine, HNSW M=16)