philiprehberger-embedding-store

In-memory vector store with multi-metric similarity search.

Installation

pip install philiprehberger-embedding-store

Usage

from philiprehberger_embedding_store import VectorStore

store = VectorStore(dimensions=1536)

# Add vectors with metadata
store.add("doc1", embedding=[0.1, 0.2, ...], metadata={"title": "First doc"})
store.add("doc2", embedding=[0.3, 0.1, ...], metadata={"title": "Second doc"})

# Search by similarity
results = store.search(query_embedding=[0.15, 0.18, ...], top_k=5)
for result in results:
    print(f"{result.id}: score={result.score:.3f}, {result.metadata}")

Distance metrics

Choose a metric per store or override per search call:

from philiprehberger_embedding_store import VectorStore

# Set default metric at store level
store = VectorStore(dimensions=128, metric="euclidean")
results = store.search(query, top_k=5)

# Override metric for a single search
results = store.search(query, top_k=5, metric="manhattan")

Supported metrics: "cosine" (default), "dot", "euclidean", "manhattan".

Metadata filtering

from philiprehberger_embedding_store import VectorStore

store = VectorStore()
store.add("d1", [1.0, 0.0], {"category": "docs", "lang": "en"})
store.add("d2", [0.9, 0.1], {"category": "code", "lang": "en"})

# Filter by single field
results = store.search(query, filter=lambda m: m["category"] == "docs")

# Filter by multiple conditions
results = store.search(
    query,
    filter=lambda m: m["category"] == "docs" and m["lang"] == "en",
)

Batch operations

from philiprehberger_embedding_store import VectorStore

store = VectorStore()

# Add many vectors at once
store.add_many([
    ("id1", [0.1, 0.2], {"label": "first"}),
    ("id2", [0.3, 0.4], {"label": "second"}),
])

# Search with multiple queries at once
all_results = store.search_many(
    [query_embedding_1, query_embedding_2],
    top_k=5,
)

Score a single entry

Use score() to compute the similarity between a stored entry and an arbitrary query vector without running a full top-k search — handy for re-ranking or one-off comparisons.

from philiprehberger_embedding_store import VectorStore

store = VectorStore(metric="cosine")
store.add("doc1", [1.0, 0.0, 0.0])

store.score("doc1", [1.0, 0.0, 0.0])              # 1.0
store.score("doc1", [0.0, 1.0, 0.0])              # ~0.0
store.score("doc1", [1.0, 1.0, 1.0], metric="dot")  # 1.0

Persistence

from philiprehberger_embedding_store import VectorStore

store = VectorStore()
store.add("doc1", [0.1, 0.2], {"title": "Example"})

# Save to disk
store.save("vectors.json")

# Load from disk
loaded = VectorStore.load("vectors.json")

Store management

from philiprehberger_embedding_store import VectorStore

store = VectorStore()
store.add("a", [1.0, 0.0])

store.remove("a")      # Remove by ID
store.clear()           # Remove all entries

Updating and clearing

from philiprehberger_embedding_store import VectorStore

store = VectorStore(dimensions=3)
store.add("a", [1.0, 0.0, 0.0], {"version": 1})

# Replace the vector in place
store.update("a", vector=[0.0, 1.0, 0.0])

# Replace the metadata (wholesale)
store.update("a", metadata={"version": 2})

# Update both at once
store.update("a", vector=[0.0, 0.0, 1.0], metadata={"version": 3})

# Remove everything but keep the dimensionality (3) and metric configuration
store.clear()
assert len(store) == 0
store.add("b", [0.1, 0.2, 0.3])  # still constrained to 3 dimensions

API

Development

pip install -e .
python -m pytest tests/ -v

Support

If you find this project useful:

⭐ Star the repo

🐛 Report issues

💡 Suggest features

❤️ Sponsor development

🌐 All Open Source Projects

💻 GitHub Profile

🔗 LinkedIn Profile

License

MIT