pulsedb 1.0.2

High-performance Python Vector Database & Memory Engine with RESP2 support.

⚡ PulseDB

An enterprise-grade, in-memory database with a native AI Vector Engine.

Built for developers who need Redis-compatible storage and lightning-fast semantic search — without running two separate systems.

---

What is PulseDB?

PulseDB is a high-performance, open-source database that combines:

• A Redis-compatible KV store — Strings, Lists, Hashes with TTL, LRU eviction, and RESP2 wire protocol
• An AI Memory Engine — HNSW-based vector search with native C++ pre-filtering callbacks
• A Python SDK — Ergonomic db.vectors.upsert() / db.vectors.search() API
• A LangChain Integration — Drop-in PulseDBVectorStore for RAG pipelines with metadata filtering

One server, one protocol, one SDK. No Pinecone. No Weaviate. No Redis Stack.

---

Features

Category	Capability
KV Store	SET, GET, DEL, EXPIRE, TTL, MSET, MGET, INCR, APPEND
Data Types	Strings · Lists (LPUSH/RPOP/LRANGE) · Hashes (HSET/HGET/HGETALL)
Vector Engine	HNSW cosine similarity, O(log N) search, dynamic resizing
Hybrid Search	Native C++ pre-filter callbacks — filter by metadata during graph traversal
Persistence	Write-Ahead Log (WAL) + JSON snapshots + HNSW binary graph snapshots
Protocol	RESP2 TCP (port 6379) — works with redis-cli, redis-py, ioredis
Cluster	Consistent hashing, multi-node routing
Auth	API Key (HTTP) + REQUIREPASS (TCP) + optional TLS/SSL
Observability	Prometheus /metrics endpoint, structured /health and /ready
LangChain	PulseDBVectorStore with similarity_search(filter={...})

---

Quickstart

1. Run the Server (Docker)

docker run -d \
  -p 6379:6379 \
  -p 8000:8000 \
  -v pulsedb_data:/app/data \
  --name pulsedb \
  ghcr.io/gkavinrajancodes/pulsedb:latest

Or use Docker Compose for a 3-node cluster:

git clone https://github.com/gkavinrajanCodes/pulseDB.git
cd pulseDB && docker-compose up --build

2. Install the SDK

pip install pulsedb

3. Use It

from pulsedb import PulseDB

db = PulseDB(host="localhost", port=6379)

# Standard KV Store
db.set("session:abc", "user_data", ttl=3600)
print(db.get("session:abc"))  # "user_data"

# AI Memory Engine — insert vectors with metadata
db.vectors.upsert("article:1", [0.12, 0.98, 0.34], metadata={"category": "sports", "year": 2024})
db.vectors.upsert("article:2", [0.91, 0.11, 0.67], metadata={"category": "tech", "year": 2023})

# Semantic similarity search — optionally filter by metadata
results = db.vectors.search([0.10, 0.95, 0.40], top_k=5, filter={"category": "sports"})
# → [{"id": "article:1", "score": 0.997}]

---

LangChain Integration

PulseDB works natively as a LangChain VectorStore, giving your RAG pipeline blazing fast retrieval with hybrid metadata filtering.

from langchain_openai import OpenAIEmbeddings
from langchain_pulsedb.vectorstore import PulseDBVectorStore

store = PulseDBVectorStore(
    embedding=OpenAIEmbeddings(),
    host="localhost",
    port=6379,
)

# Ingest documents — metadata is automatically stored for hybrid filtering
store.add_texts(
    texts=["PulseDB is fast", "Redis is popular", "Pinecone is expensive"],
    metadatas=[{"source": "blog"}, {"source": "wiki"}, {"source": "review"}]
)

# Hybrid search — find similar docs but only from the blog source
docs = store.similarity_search("fast database", k=2, filter={"source": "blog"})

---

How the AI Memory Engine Works

Standard vector databases do post-filtering: search all vectors, get K results, then throw away the ones that don't match the filter. This degrades accuracy.

PulseDB does true pre-filtering using native hnswlib C++ filter callbacks. The filter function is evaluated inside the graph traversal — so the C++ engine skips disqualified nodes entirely before scoring them.

Query Vector → HNSW Graph Traversal → [Filter Callback runs on every node visited]
                                        ↓ Pass → included in result set
                                        ↓ Fail → skipped immediately
                                       Top-K results returned

This means your effective top_k is always accurate, even with highly restrictive filters.

---

Architecture

graph TD
    Client["Client (SDK / redis-cli)"] -->|RESP2 Binary Protocol| TCP["asyncio TCP Server :6379"]
    Client -->|HTTP REST| HTTP["FastAPI Gateway :8000"]
    TCP --> Router["Command Router"]
    HTTP --> Router
    Router --> KV["16-Shard KV Store (LRU + TTL)"]
    Router --> VE["AI Vector Engine (hnswlib HNSW)"]
    Router --> DT["Data Types (Lists, Hashes)"]
    Router --> PS["Pub/Sub Engine"]
    KV --> WAL["Write-Ahead Log"]
    VE --> Snap["HNSW Binary Snapshot"]
    WAL --> Snap

---

Run Locally (From Source)

# 1. Clone and install
git clone https://github.com/gkavinrajanCodes/pulseDB.git
cd pulseDB
python3.10 -m venv workenv && source workenv/bin/activate
pip install -r requirements.txt

# 2. Start the server
NODE_ID=node1 CLUSTER_NODES=node1 uvicorn server.main:app --host 0.0.0.0 --port 8000

# 3. Install the SDK (in another terminal)
pip install -e sdk/

---

Contributing

1. Fork the repository
2. Create a feature branch: git checkout -b feature/sorted-sets
3. Commit your changes: git commit -m "feat: add ZADD/ZRANGE sorted set commands"
4. Push: git push origin feature/sorted-sets
5. Open a Pull Request

All PRs are validated against our CI matrix (Python 3.10, 3.11, 3.12 with flake8, mypy, and pytest).

---

License

Distributed under the Business Source License (BSL 1.1). See LICENSE for details.