To combine metadata filters with vector similarity search, you typically do one of these:

1) Filter first, then search within the filtered set

Use metadata conditions to narrow the candidate documents, then run vector similarity on that subset.

Example pattern:

• Metadata: category = "legal" and year >= 2023
• Vector search: find the top 10 most similar embeddings among those documents

This is common when your search system supports a where / filter clause alongside vector search.

Pseudo-query:

{
  "vector": [0.12, 0.98, ...],
  "top_k": 10,
  "filter": {
    "category": "legal",
    "year": { "$gte": 2023 }
  }
}

2) Search first, then filter results

Run vector search over all items, then remove results that don’t match metadata.

This is simpler, but can hurt recall:

• If the top similar items don’t satisfy the filter, they get discarded
• You may end up with fewer than top_k results

3) Hybrid/combined retrieval

Some systems combine:

• metadata filtering
• vector similarity
• sometimes keyword/BM25 search

This is often the best approach for production search.

---

Practical considerations

If your filter is restrictive

Example:

• tenant_id = 123
• language = "en"
• status = "published"

Filtering first is usually better, because it reduces the search space and improves relevance within that slice.

If your vector DB supports indexes on metadata

Make sure metadata fields used in filters are indexed if possible. That improves performance.

If your DB uses approximate nearest neighbors (ANN)

Filtering may be implemented as:

• pre-filtering: search only within filtered candidates
• post-filtering: search broadly, then apply filters
• integrated filtering: the engine handles both together

Integrated filtering is usually most efficient.

---

Example with a typical vector database API

results = collection.query(
    vector=query_embedding,
    top_k=5,
    filter={
        "doc_type": "report",
        "published": True
    }
)

This returns the 5 nearest vectors among documents where:

• doc_type == "report"
• published == True

---

Rule of thumb

• Use metadata filters for hard constraints: tenant, permissions, language, date range, category
• Use vector similarity for semantic ranking
• Prefer a database/query engine that supports both in one query

If you tell me which vector database or search stack you’re using (Pinecone, Weaviate, Chroma, Milvus, Elasticsearch, pgvector, etc.), I can show the exact syntax.