> ## Documentation Index
> Fetch the complete documentation index at: https://memwirelabs.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# FastEmbed

> Run local dense and sparse embeddings with Qdrant's FastEmbed library — no external API required.

[FastEmbed](https://github.com/qdrant/fastembed) is MemWire's default embedding backend. Models run entirely on-device — no OpenAI key, no network calls, no per-token cost.

Two model types are used:

| Type             | Purpose                                                          |
| ---------------- | ---------------------------------------------------------------- |
| **Dense model**  | Semantic similarity — powers memory search and recall            |
| **Sparse model** | Keyword matching — used alongside dense vectors in hybrid search |

***

## Default models

No configuration needed to get started — MemWire ships with sensible defaults.

```python theme={null}
from memwire import MemWire, MemWireConfig

config = MemWireConfig(
    model_name="sentence-transformers/all-MiniLM-L6-v2",
    embedding_dim=384,
    sparse_model_name="prithivida/Splade_PP_en_v1",
)
memory = MemWire(config=config)
```

| Model                                    | Type   | Dimensions | Notes                                                      |
| ---------------------------------------- | ------ | ---------- | ---------------------------------------------------------- |
| `sentence-transformers/all-MiniLM-L6-v2` | Dense  | 384        | Fast, multilingual-friendly, good general-purpose baseline |
| `prithivida/Splade_PP_en_v1`             | Sparse | —          | SPLADE sparse model for hybrid BM25+vector search          |

***

## Changing the dense model

You can swap the dense model for any [FastEmbed-compatible model](https://qdrant.github.io/fastembed/examples/Supported_Models/). Make sure `embedding_dim` matches the model's output dimension.

```python theme={null}
# Smaller — faster inference, slightly lower accuracy
config = MemWireConfig(
    model_name="BAAI/bge-small-en-v1.5",
    embedding_dim=384,
)
```

```python theme={null}
# Balanced — good quality, moderate size
config = MemWireConfig(
    model_name="BAAI/bge-base-en-v1.5",
    embedding_dim=768,
)
```

```python theme={null}
# Larger — best quality, slower inference
config = MemWireConfig(
    model_name="BAAI/bge-large-en-v1.5",
    embedding_dim=1024,
)
```

<Warning>
  If you change the embedding model after data has already been stored, existing vectors will be incompatible with new embeddings. Reset your Qdrant storage when switching models.
</Warning>

***

## Disabling hybrid search

If you want dense-only retrieval and don't need the sparse model to load:

```python theme={null}
config = MemWireConfig(
    use_hybrid_search=False,
)
```

***

## Enabling reranking

A cross-encoder reranker re-scores the top candidates after initial retrieval. This improves precision at the cost of slightly higher latency.

```python theme={null}
config = MemWireConfig(
    use_reranking=True,
    reranker_model_name="Xenova/ms-marco-MiniLM-L-6-v2",
)
```

***

## Configuration reference

| Parameter                 | Default                                  | Description                                                               |
| ------------------------- | ---------------------------------------- | ------------------------------------------------------------------------- |
| `model_name`              | `sentence-transformers/all-MiniLM-L6-v2` | Dense embedding model. Must be FastEmbed-compatible.                      |
| `embedding_dim`           | `384`                                    | Output dimension of the dense model. Must match the model.                |
| `sparse_model_name`       | `prithivida/Splade_PP_en_v1`             | Sparse model for hybrid BM25+vector search.                               |
| `use_hybrid_search`       | `True`                                   | Combine dense and sparse vectors. Set to `False` for dense-only search.   |
| `use_reranking`           | `False`                                  | Apply a cross-encoder reranker to re-score top candidates.                |
| `reranker_model_name`     | `Xenova/ms-marco-MiniLM-L-6-v2`          | Cross-encoder reranker model (used when `use_reranking=True`).            |
| `embedding_cache_maxsize` | `10000`                                  | LRU cache size for embedding vectors. Increase for large-scale workloads. |