Sentence Transformer

Local reranking with HuggingFace cross-encoder models

Sentence Transformer reranker provides local reranking using HuggingFace cross-encoder models, perfect for privacy-focused deployments where you want to keep data on-premises.

Models

Any HuggingFace cross-encoder model can be used. Popular choices include:

• cross-encoder/ms-marco-MiniLM-L-6-v2: Default, good balance of speed and accuracy
• cross-encoder/ms-marco-TinyBERT-L-2-v2: Fastest, smaller model size
• cross-encoder/ms-marco-electra-base: Higher accuracy, larger model
• cross-encoder/stsb-distilroberta-base: Good for semantic similarity tasks

Installation

pip install sentence-transformers

Configuration

Python

from mem0 import Memory

config = {
    "vector_store": {
        "provider": "chroma",
        "config": {
            "collection_name": "my_memories",
            "path": "./chroma_db"
        }
    },
    "llm": {
        "provider": "openai",
        "config": {
            "model": "gpt-4o-mini"
        }
    },
    "rerank": {
        "provider": "sentence_transformer",
        "config": {
            "model": "cross-encoder/ms-marco-MiniLM-L-6-v2",
            "device": "cpu",  # or "cuda" for GPU
            "batch_size": 32,
            "show_progress_bar": False,
            "top_k": 5
        }
    }
}

memory = Memory.from_config(config)

TypeScript (self-hosted)

The TypeScript OSS SDK (mem0ai/oss) runs this reranker locally with Transformers.js. Because it executes ONNX weights, the default model is the ONNX mirror of the Python default: Xenova/ms-marco-MiniLM-L-6-v2. Point model at any ONNX-exported cross-encoder on the Hub (a raw cross-encoder/... PyTorch checkpoint will not load in this runtime).

pnpm add @huggingface/transformers

import { Memory } from "mem0ai/oss";

const memory = new Memory({
  reranker: {
    provider: "sentence_transformer",
    config: {
      // model: "Xenova/ms-marco-MiniLM-L-6-v2", // default (ONNX)
      device: "cpu", // "cpu" | "wasm" | "webgpu"
      maxLength: 512, // max tokens per query-document pair
      normalize: true, // sigmoid-normalize logits to [0, 1] (default)
      topK: 5,
    },
  },
});

const results = await memory.search("What books does the user like?", {
  filters: { userId: "charlie" },
  rerank: true,
});

batchSize and showProgressBar are accepted for parity with the Python SDK but are no-ops in the TypeScript runtime, because a search reranks a small candidate set in a single in-process forward pass. The model downloads once and is cached in-process.

GPU Acceleration

For better performance, use GPU acceleration:

Python

config = {
    "rerank": {
        "provider": "sentence_transformer",
        "config": {
            "model": "cross-encoder/ms-marco-MiniLM-L-6-v2",
            "device": "cuda",  # Use GPU
            "batch_size": 64   # high batch size for high memory GPUs
        }
    }
}

Usage Example

Python

from mem0 import Memory

# Initialize memory with local reranker
config = {
    "vector_store": {"provider": "chroma"},
    "llm": {"provider": "openai", "config": {"model": "gpt-4o-mini"}},
    "rerank": {
        "provider": "sentence_transformer",
        "config": {
            "model": "cross-encoder/ms-marco-MiniLM-L-6-v2",
            "device": "cpu"
        }
    }
}

memory = Memory.from_config(config)

# Add memories
messages = [
    {"role": "user", "content": "I love reading science fiction novels"},
    {"role": "user", "content": "My favorite author is Isaac Asimov"},
    {"role": "user", "content": "I also enjoy watching sci-fi movies"}
]

memory.add(messages, user_id="charlie")

# Search with local reranking
results = memory.search("What books does the user like?", filters={"user_id": "charlie"})

for result in results['results']:
    print(f"Memory: {result['memory']}")
    print(f"Vector Score: {result['score']:.3f}")
    print(f"Rerank Score: {result['rerank_score']:.3f}")
    print()

Custom Models

You can use any HuggingFace cross-encoder model:

Python

# Using a different model
config = {
    "rerank": {
        "provider": "sentence_transformer", 
        "config": {
            "model": "cross-encoder/stsb-distilroberta-base",
            "device": "cpu"
        }
    }
}

Configuration Parameters

Parameter	Description	Type	Default
model	HuggingFace cross-encoder model name	str	"cross-encoder/ms-marco-MiniLM-L-6-v2"
device	Device to run model on (cpu, cuda, etc.)	str	None
batch_size	Batch size for processing documents	int	32
show_progress_bar	Show progress bar during processing	bool	False
top_k	Maximum documents to return	int	None

Advantages

• Privacy: Complete local processing, no external API calls
• Cost: No per-token charges after initial model download
• Customization: Use any HuggingFace cross-encoder model
• Offline: Works without internet connection after model download

Performance Considerations

• First Run: Model download may take time initially
• Memory Usage: Models require GPU/CPU memory
• Batch Size: Optimize batch size based on available memory
• Device: GPU acceleration significantly improves speed

Best Practices

1. Model Selection: Choose model based on accuracy vs speed requirements
2. Device Management: Use GPU when available for better performance
3. Batch Processing: Process multiple documents together for efficiency
4. Memory Monitoring: Monitor system memory usage with larger models