Vector Databases

Storing and searching embeddings for semantic retrieval and RAG.

Quick Start

# pgvector on Postgres
docker run -e POSTGRES_PASSWORD=secret -p 5432:5432 -d pgvector/pgvector:pg17
# or Chroma:
pip install chromadb

When to Use

• Semantic search over unstructured data
• RAG retrieval backends
• Recommendation by similarity
• Deduplication and clustering of texts

Best Practices

Embeddings

• Use retrieval-tuned embedding models
• Normalize vectors for cosine similarity
• Choose dimension balanced with storage and speed
• Embed at indexing time and query time consistently

Indexes

• Prefer HNSW for balanced speed/recall
• Set `m` (connections) and `ef_construction` for build quality
• Use IVF for very large collections with disk budget
• Match the index to the metric (cosine/euclidean)

Filtering & Hybrid

• Combine vector search with metadata filters
• Support hybrid retrieval (vector + keyword) when needed
• Re-rank top-k with a cross-encoder for quality
• Return scores for confidence thresholds

Operations

• Set replication and backup for durability
• Monitor recall, latency, and index size
• Re-index when the embedding model changes
• Version the embedding model with the data

Dependencies

# pgvector
docker run -e POSTGRES_PASSWORD=secret -p 5432:5432 -d pgvector/pgvector:pg17
# or Python-native
pip install chromadb qdrant-client

Examples

-- pgvector setup and query
CREATE EXTENSION IF NOT EXISTS vector;

CREATE TABLE documents (
  id bigserial PRIMARY KEY,
  content text,
  embedding vector(384)
);

CREATE INDEX ON documents USING hnsw (embedding vector_cosine_ops);

-- Similarity search with filter
SELECT id, content, 1 - (embedding <=> $1) AS similarity
FROM documents
WHERE category = 'docs'
ORDER BY embedding <=> $1
LIMIT 5;

import chromadb

client = chromadb.PersistentClient(path="./db")
col = client.get_or_create_collection("docs")

col.add(
    ids=["1", "2"],
    embeddings=[[0.1] * 384, [0.9] * 384],
    documents=["Alpha docs", "Beta docs"],
    metadatas=[{"cat": "guide"}, {"cat": "api"}],
)

res = col.query(query_embeddings=[[0.1] * 384], n_results=2, where={"cat": "guide"})
print(res["documents"])

# Embed + store with sentence-transformers
from sentence_transformers import SentenceTransformer

model = SentenceTransformer("all-MiniLM-L6-v2")
vec = model.encode("query text").tolist()
col.add(ids=[id], embeddings=[vec], documents=[text])

Step-by-Step

1. Choose a store: pgvector, Chroma, Qdrant, or managed services.

2. Pick an embedding model and vector dimension.

3. Index your corpus in batches with metadata.

4. Create an HNSW index tuned to your recall/latency goals.

5. Query with filters and hybrid options.

6. Re-rank results for quality where it matters.

7. Set backups, replication, and monitoring.

8. Re-index when you upgrade the embedding model.

Validation

1. Recall@k meets the target on a labeled set

2. Query latency is within budget at scale

3. Metadata filters return correct subsets

4. Scores are meaningful and thresholdable

5. Backups restore successfully

Troubleshooting

• Low recall: tune HNSW params or switch index type.
• Slow queries: reduce dimensions, add filters, or scale shards.
• Stale results: re-index after embedding model changes.