Prompt
How do I search embeddings with metadata filters like tenant and date?
Latest observation
Oct 28, 2026
OpenAI APIWeb search: off
To search embeddings with metadata filters like tenant and date, you typically:
- Store metadata alongside each embedding
- Build a query embedding for the user’s search text
- Apply metadata filters during retrieval
- Rank results by vector similarity within the filtered subset
Example pattern
1) When indexing
Store each record like:
{
"id": "doc_123",
"text": "Invoice payment terms ...",
"embedding": [0.12, -0.03, ...],
"metadata": {
"tenant": "acme",
"date": "2026-10-01",
"doc_type": "invoice"
}
}
2) At query time
You embed the search query:
query_embedding = embed("payment terms for overdue invoices")
Then search with filters:
results = vector_db.search(
embedding=query_embedding,
top_k=10,
filter={
"tenant": "acme",
"date": {"$gte": "2026-01-01", "$lt": "2027-01-01"}
}
)
Common filter logic
- Exact match for tenant:
tenant == "acme"
- Range filter for date:
date >= start_datedate < end_date
Example in different vector stores
Pinecone
index.query(
vector=query_embedding,
top_k=10,
filter={
"tenant": {"$eq": "acme"},
"date": {"$gte": "2026-01-01", "$lt": "2027-01-01"}
}
)
Weaviate
where_filter = {
"operator": "And",
"operands": [
{"path": ["tenant"], "operator": "Equal", "valueString": "acme"},
{"path": ["date"], "operator": "GreaterThanEqual", "valueDate": "2026-01-01T00:00:00Z"}
]
}
Elasticsearch / OpenSearch
Use:
knnorscript_scorefor vector similaritybool.filterfor metadata constraints
{
"query": {
"bool": {
"filter": [
{ "term": { "tenant": "acme" } },
{ "range": { "date": { "gte": "2026-01-01", "lt": "2027-01-01" } } }
],
"must": [
{
"knn": {
"embedding": {
"vector": [0.12, -0.03],
"k": 10
}
}
}
]
}
}
}
Important best practices
- Use structured date formats like ISO-8601 (
YYYY-MM-DDor full timestamp) - Index filterable fields (
tenant,date,doc_type) separately from text - Apply tenant isolation first to avoid cross-tenant leakage
- Normalize time zones if filtering by timestamps
- If your DB doesn’t support pre-filtering, do post-filtering but note it can hurt relevance/recall
If your vector DB doesn’t support metadata filters
You can:
- Retrieve top
Nnearest neighbors - Filter those results by metadata
- If too few remain, increase
Nand retry
That’s less efficient, but works.
If you tell me which vector database you’re using (Pinecone, Weaviate, pgvector, FAISS, Chroma, Elasticsearch/OpenSearch, etc.), I can give you the exact query syntax.