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

# Hybrid Search

> Combine dense semantic vectors and sparse SPLADE keyword vectors for higher-recall retrieval.

## How it works

Pure semantic (dense) search misses exact-match queries like product codes, names, or technical terms. Pure keyword (sparse) search misses paraphrases and synonyms. MemWire combines both using **Qdrant's hybrid search** with reciprocal rank fusion:

* **Dense vectors** — sentence embeddings from `TextEmbedding` (FastEmbed), capturing semantic meaning
* **Sparse vectors** — SPLADE token weights from `SparseTextEmbedding`, capturing exact lexical signal

Both vectors are computed for every stored memory and every query. Results are merged by Qdrant before being returned to MemWire.

***

## Enabling hybrid search

Hybrid search is **on by default**. No extra configuration needed:

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

config = MemWireConfig(qdrant_path="./memwire_data")   # use_hybrid_search=True by default
memory = MemWire(config=config)
```

***

## Disabling hybrid search

If you want dense-only retrieval (faster, lower memory):

```python theme={null}
config = MemWireConfig(
    qdrant_path="./memwire_data",
    use_hybrid_search=False,
)
```

<Warning>
  Disabling hybrid search means the sparse model is never loaded. This saves \~200 MB of RAM but reduces retrieval quality for exact-match queries.
</Warning>

***

## Adding a cross-encoder reranker

For the highest retrieval quality, combine hybrid search with a cross-encoder reranker. The reranker re-scores the top candidates using full query-document attention:

```python theme={null}
config = MemWireConfig(
    qdrant_path="./memwire_data",
    use_hybrid_search=True,
    use_reranking=True,
    reranker_model_name="Xenova/ms-marco-MiniLM-L-6-v2",  # default
)
memory = MemWire(config=config)

results = memory.search("deadline for the project", user_id="alice", top_k=5)
for record, score in results:
    print(f"[{score:.3f}] {record.content}")
```

The reranker is **lazy-loaded** — it is only downloaded and initialised on the first `search()` call.

***

## Models

| Role             | Default model                            | Notes                   |
| ---------------- | ---------------------------------------- | ----------------------- |
| Dense embedding  | `sentence-transformers/all-MiniLM-L6-v2` | 384-dim, fast           |
| Sparse embedding | `prithivida/Splade_PP_en_v1`             | SPLADE++                |
| Reranker         | `Xenova/ms-marco-MiniLM-L-6-v2`          | Cross-encoder, optional |

Swap any model via `MemWireConfig`:

```python theme={null}
config = MemWireConfig(
    model_name="BAAI/bge-small-en-v1.5",
    sparse_model_name="prithivida/Splade_PP_en_v1",
    reranker_model_name="Xenova/ms-marco-MiniLM-L-6-v2",
    embedding_dim=384,
)
```

***

## Configuration reference

| Parameter                 | Default                                  | Description                                     |
| ------------------------- | ---------------------------------------- | ----------------------------------------------- |
| `use_hybrid_search`       | `True`                                   | Combine dense and sparse vectors for retrieval. |
| `use_reranking`           | `False`                                  | Apply a cross-encoder reranker to top results.  |
| `model_name`              | `sentence-transformers/all-MiniLM-L6-v2` | Dense embedding model.                          |
| `sparse_model_name`       | `prithivida/Splade_PP_en_v1`             | Sparse (SPLADE) embedding model.                |
| `reranker_model_name`     | `Xenova/ms-marco-MiniLM-L-6-v2`          | Cross-encoder reranker model.                   |
| `embedding_dim`           | `384`                                    | Dimension of the dense embedding.               |
| `embedding_cache_maxsize` | `10000`                                  | LRU cache size for embedding vectors.           |