how to index embeddings in qdrant with filters

In Qdrant, you usually don’t “index embeddings with filters” as one combined thing. Instead, you:

1. Store embeddings in a vector field
2. Store metadata in payload fields
3. Create payload indexes on the fields you want to filter by
4. Search with a vector + filter at query time

1) Create a collection with vectors

Example using sentence-transformers style vectors:

from qdrant_client import QdrantClient
from qdrant_client.models import VectorParams, Distance

client = QdrantClient(url="http://localhost:6333")

client.create_collection(
    collection_name="docs",
    vectors_config=VectorParams(size=384, distance=Distance.COSINE),
)

2) Insert vectors with payload metadata

Put your filterable fields in payload:

client.upsert(
    collection_name="docs",
    points=[
        {
            "id": 1,
            "vector": [0.1] * 384,
            "payload": {"category": "news", "lang": "en", "year": 2024},
        },
        {
            "id": 2,
            "vector": [0.2] * 384,
            "payload": {"category": "blog", "lang": "en", "year": 2023},
        },
    ],
)

3) Create payload indexes for filterable fields

This helps Qdrant filter faster:

from qdrant_client.models import PayloadSchemaType

client.create_payload_index(
    collection_name="docs",
    field_name="category",
    field_schema=PayloadSchemaType.KEYWORD,
)

client.create_payload_index(
    collection_name="docs",
    field_name="lang",
    field_schema=PayloadSchemaType.KEYWORD,
)

client.create_payload_index(
    collection_name="docs",
    field_name="year",
    field_schema=PayloadSchemaType.INTEGER,
)

Common payload index types

• KEYWORD for exact-match strings
• INTEGER / FLOAT for numeric ranges
• BOOL
• DATETIME
• GEO

4) Search with a vector and a filter

Example: only search English news from 2024:

from qdrant_client.models import Filter, FieldCondition, MatchValue, Range, SearchParams

query_vector = [0.1] * 384

results = client.search(
    collection_name="docs",
    query_vector=query_vector,
    limit=5,
    query_filter=Filter(
        must=[
            FieldCondition(key="category", match=MatchValue(value="news")),
            FieldCondition(key="lang", match=MatchValue(value="en")),
            FieldCondition(key="year", range=Range(gte=2024)),
        ]
    ),
)

Important note

Qdrant uses the vector index for similarity search and the payload index for filtering.
So “index embeddings with filters” really means:

• Vector index: for nearest-neighbor search
• Payload index: for metadata filters

Optional: filter before or after search?

Qdrant can combine them efficiently. Depending on the dataset and filter selectivity, it may:

• filter first, then search a smaller subset
• search candidates, then apply filter
• use optimized hybrid logic internally

Best practices

• Index fields you filter on often
• Use exact types (KEYWORD, INTEGER, etc.)
• Avoid storing filterable data only inside the text blob
• Keep payload small and structured

If you want, I can also show:

• a REST API example
• a LangChain example
• how to do hybrid search + filters in Qdrant