GitHub - imesde/imesde: The lightning-fast ephemeral vector engine for real-time AI context

⚡️ imesde

The Real-Time Vector Database. "Semantic search at the speed of a pipe. No disk, no lag, just context."

Welcome to imesde, the In-Memory Streaming Data Engine. We are defining a new category in the data landscape: the IMESDE.

🎯 Why imesde?

Traditional vector databases are built for persistence and long-term storage. imesde is built for speed and ephemerality:

• Zero-Disk Dependency: Pure RAM operation. Old data flows out as new data flows in. No GC, no fragmentation.

• Real-Time RAG Engine: Enables the "Infinite Window". Feed LLMs with live context (logs, tickers, chats) with zero indexing latency and automatic "forgetting" of stale data.

• O(1) Instant Anomaly Detection: Built-in support for Sliding Window Centroids. Automatically detect "unseen" anomalies during ingestion by measuring semantic distance from the mathematical mean in real-time.

• Local-First Privacy: In-process vectorization (ONNX) and storage. Data never leaves your machine.

Use Cases

🚀 imesde Performance Benchmark

💻 System: Apple M4 (Darwin 24.6.0) - 16GB RAM
📊 Dataset: 5000 records
⚙️ Config: 16 Shards × 1024 Record Size
🧠 Model: bge-small-en-v1.5 int8

🏆 imesde vs Qdrant (Pure Engine)

A pure engine-to-engine comparison (excluding AI embedding time) between imesde and Qdrant (In-Memory mode). This benchmark measures the raw speed of the underlying Rust search kernels.

Test Setup: 20,000 records, 384 dimensions, Apple M4 (Darwin 24.6.0).
Source Code: benchmark_vs_qdrant.py

⚖️ When to use which?

🚀 Where imesde wins:

• 🤖 AI Agent Memory (Short-Term): Ideal for agents that need to recall "what happened in the last 10 minutes." The dataset is small but highly volatile. imesde provides instant speed, whereas a full Vector DB would be an unnecessary waste of resources for such a small volume.
• 🌊 High-Frequency Streams: Critical for system logs, financial feeds, or IoT sensors (~10,000 events/sec) requiring real-time anomaly detection. imesde ingests and searches without lag, avoiding the performance hit caused by constant re-indexing in traditional DBs.

🏆 Recommended Models

General Recommendation: For the best balance of speed and efficiency on CPUs, we generally recommend using Int8 quantized models.

🧠 Why CPU-First?

imesde is intentionally architected to run on CPUs, not GPUs. While GPUs offer high throughput for massive batch training, they introduce latency (PCIe data transfer) and operational complexity that contradict the goal of a lightweight, real-time streaming engine.

The Strategy:

1. Zero-Latency: No data transfer between RAM and VRAM.
2. Quantization is King: Modern CPUs with AVX2/NEON/AMX process Int8 Quantized models at monstrous speeds.
3. Result: GPU-class inference for streaming data with significantly lower operational complexity.

Need higher precision? If absolute semantic accuracy > latency, you can simply drop in a standard Float32 model (e.g., bge-large, e5-mistral). imesde works with any ONNX model out of the box.

🚀 Quick Start

# Build Rust binary
cargo build --release

🐍 Python Usage

For a detailed guide on using imesde with Python, see the Python Documentation.

📖 Documentation & Use Cases

• Real-Time RAG (The Infinite Window): How to use imesde for live context retrieval.
• AI & LLM Integration: Examples with Ollama and OpenAI.
• 🛰️ Showcase: Semantic Radar: A full-scale stress test monitoring global aviation data with AI reasoning.

  python bindings/python/examples/semantic_radar.py

You can install imesde directly via pip:

Example

from imesde import PyImesde

# Initialize with O(1) centroid tracking enabled
engine = PyImesde("model/model.onnx", "model/tokenizer.json", track_centroid=True)

# Single ingestion returns an instant anomaly score (similarity to mean)
score = engine.ingest("Real-time log data flow")
if score < 0.45:
    print(f"🚨 Instant anomaly detected: {score}")

# High-performance batch ingestion (Parallelized)
logs = ["User login at 10:00", "DB Query took 500ms", "Connection reset"]
scores = engine.ingest_batch(logs)

# Search the circular buffer
results = engine.search("database issues", k=5)
for text, score in results:
    print(f"[{score:.4f}] {text}")

# Get the mathematical mean (O(1) retrieval)
centroid = engine.get_centroid()

Note: imesde requires an ONNX model and its tokenizer. You can export these from Hugging Face using optimum-cli or sentence-transformers. Place them in the model/ directory as model.onnx and tokenizer.json.

MIT Licensed. Built for the speed of thought.