# DenSpa
**DenSpa** is an open-source package designed for hybrid search, enabling seamless integration into Retrieval-Augmented Generation (RAG) frameworks. The package combines **dense** and **sparse vector embeddings** to perform efficient searches on document corpora.
- **Dense-vector-based search** leverages the FAISS vector database to manage and query collections.
- **Sparse-vector-based search** utilizes a custom implementation of BM25, enhanced with pre-processing techniques like stemming to optimize the index.
## Installation
To get started, clone this repository and install the dependencies or use pip:
```bash
pip install DenSpa
```
## Quick Start
### Initializing the Vector Search Engine
You can easily initialize the vector search engine using the following code:
```python
from denspa import VectorSearch
from langchain.embeddings import HuggingFaceEmbeddings
import os
embedding_function = HuggingFaceEmbeddings(model_name="sentence-transformers/all-mpnet-base-v2")
INDEX_PATH = "database/index"
if not os.path.exists(INDEX_PATH):
os.makedirs(INDEX_PATH)
vecsea = VectorSearch(
folder_path=INDEX_PATH,
index_name="denspa",
embedding_function=embedding_function,
bm25_options={"k1": 1.25, "b": 0}
)
```
### Indexing Documents
DenSpa supports indexing documents in both **English** and **German**. Documents can be added to the search engine like this:
```python
from langchain.docstore.document import Document
documents = [Document(page_content="There are many variations...", metadata={"source": "lecture.pdf"})]
# Indexing with FAISS
vecsea.add_documents(documents, lang="en", engine="faiss")
# Indexing with BM25
vecsea.add_documents(documents, lang="en", engine="bm25")
# Save the index locally
vecsea.save_local()
```
### Deleting Indexed Documents
To remove a specific document from the index, use the `removeByMetadata` function:
```python
vecsea.removeByMetadata({"source": "lecture.pdf"})
vecsea.save_local()
```
### Deleting Indexes
To remove the indexes, use the `delete_local` function:
```python
vecsea.delete_local()
```
## Search Methods
DenSpa currently supports **three search methods**:
1. **FAISS**: Semantic search that uses dense vectors for similarity.
2. **BM25**: Keyword-based search leveraging sparse vectors.
3. **Hybrid Search**: A cascade method combining FAISS and BM25. Hybrid search first retrieves the top results using FAISS (high recall) and then applies BM25 on the top-3*k results to refine the selection for the final top-k results (higher precision).
Example usage:
```python
results = vecsea.similarity_search_with_score(
query="Quantum mechanics",
k=3,
method="bm25",
lang="en"
)
```
## Features
- **Dense and Sparse Search**: Utilize semantic embeddings and keyword-based indexing for versatile search capabilities.
- **Hybrid Search Strategy**: Combine the strengths of both FAISS and BM25 for balanced recall and precision.
- **Customizable**: Easily configure embeddings, BM25 parameters, and storage paths.
- **Language Support**: Works with English and German document corpora.
## Contributions
Contributions are welcome! Please feel free to open an issue or submit a pull request if you have suggestions or improvements.
Raw data
{
"_id": null,
"home_page": null,
"name": "DenSpa",
"maintainer": null,
"docs_url": null,
"requires_python": ">=3.10",
"maintainer_email": null,
"keywords": "hybrid search, semantic search, keyword search, vector search, FAISS BM25, LangChain",
"author": "Ivan Kankeu",
"author_email": null,
"download_url": "https://files.pythonhosted.org/packages/52/b1/6c67864c8bc5b56a15718c320727e90ffcf7bc4d05c9d3cb28c99979666e/denspa-0.0.3.tar.gz",
"platform": null,
"description": "# DenSpa\n\n**DenSpa** is an open-source package designed for hybrid search, enabling seamless integration into Retrieval-Augmented Generation (RAG) frameworks. The package combines **dense** and **sparse vector embeddings** to perform efficient searches on document corpora. \n\n- **Dense-vector-based search** leverages the FAISS vector database to manage and query collections. \n- **Sparse-vector-based search** utilizes a custom implementation of BM25, enhanced with pre-processing techniques like stemming to optimize the index. \n\n## Installation\n\nTo get started, clone this repository and install the dependencies or use pip: \n```bash\npip install DenSpa\n```\n\n## Quick Start\n\n### Initializing the Vector Search Engine\nYou can easily initialize the vector search engine using the following code: \n\n```python\nfrom denspa import VectorSearch\nfrom langchain.embeddings import HuggingFaceEmbeddings\nimport os\n\nembedding_function = HuggingFaceEmbeddings(model_name=\"sentence-transformers/all-mpnet-base-v2\")\n\nINDEX_PATH = \"database/index\"\nif not os.path.exists(INDEX_PATH):\n os.makedirs(INDEX_PATH)\n\nvecsea = VectorSearch(\n folder_path=INDEX_PATH,\n index_name=\"denspa\",\n embedding_function=embedding_function,\n bm25_options={\"k1\": 1.25, \"b\": 0}\n)\n```\n\n### Indexing Documents\nDenSpa supports indexing documents in both **English** and **German**. Documents can be added to the search engine like this:\n\n```python\nfrom langchain.docstore.document import Document\n\ndocuments = [Document(page_content=\"There are many variations...\", metadata={\"source\": \"lecture.pdf\"})]\n\n# Indexing with FAISS\nvecsea.add_documents(documents, lang=\"en\", engine=\"faiss\")\n\n# Indexing with BM25\nvecsea.add_documents(documents, lang=\"en\", engine=\"bm25\")\n\n# Save the index locally\nvecsea.save_local()\n```\n\n### Deleting Indexed Documents\nTo remove a specific document from the index, use the `removeByMetadata` function:\n\n```python\nvecsea.removeByMetadata({\"source\": \"lecture.pdf\"})\nvecsea.save_local()\n```\n\n### Deleting Indexes\nTo remove the indexes, use the `delete_local` function:\n\n```python\nvecsea.delete_local()\n```\n\n## Search Methods\n\nDenSpa currently supports **three search methods**: \n\n1. **FAISS**: Semantic search that uses dense vectors for similarity. \n2. **BM25**: Keyword-based search leveraging sparse vectors. \n3. **Hybrid Search**: A cascade method combining FAISS and BM25. Hybrid search first retrieves the top results using FAISS (high recall) and then applies BM25 on the top-3*k results to refine the selection for the final top-k results (higher precision). \n\nExample usage: \n```python\nresults = vecsea.similarity_search_with_score(\n query=\"Quantum mechanics\",\n k=3,\n method=\"bm25\",\n lang=\"en\"\n)\n```\n\n## Features\n- **Dense and Sparse Search**: Utilize semantic embeddings and keyword-based indexing for versatile search capabilities. \n- **Hybrid Search Strategy**: Combine the strengths of both FAISS and BM25 for balanced recall and precision. \n- **Customizable**: Easily configure embeddings, BM25 parameters, and storage paths. \n- **Language Support**: Works with English and German document corpora. \n\n## Contributions\nContributions are welcome! Please feel free to open an issue or submit a pull request if you have suggestions or improvements.\n",
"bugtrack_url": null,
"license": null,
"summary": "DenSpa is an open-source package designed for hybrid search, enabling seamless integration into RAG frameworks.",
"version": "0.0.3",
"project_urls": {
"Homepage": "https://github.com/Kankeu/denspa",
"Issues": "https://github.com/Kankeu/denspa/issues"
},
"split_keywords": [
"hybrid search",
" semantic search",
" keyword search",
" vector search",
" faiss bm25",
" langchain"
],
"urls": [
{
"comment_text": null,
"digests": {
"blake2b_256": "d9b0dad7ab67df467855bce74180d7efb23c590c1bc31d35e968e076902b78d8",
"md5": "27ab3a59bd4be03544d85ca0322e96d3",
"sha256": "fd34f4922bfbbf5f7579bb81b0187bdde72f4c9640e79cf4a0e15e1371a7f35a"
},
"downloads": -1,
"filename": "denspa-0.0.3-py3-none-any.whl",
"has_sig": false,
"md5_digest": "27ab3a59bd4be03544d85ca0322e96d3",
"packagetype": "bdist_wheel",
"python_version": "py3",
"requires_python": ">=3.10",
"size": 10903,
"upload_time": "2025-03-26T16:41:46",
"upload_time_iso_8601": "2025-03-26T16:41:46.219264Z",
"url": "https://files.pythonhosted.org/packages/d9/b0/dad7ab67df467855bce74180d7efb23c590c1bc31d35e968e076902b78d8/denspa-0.0.3-py3-none-any.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": null,
"digests": {
"blake2b_256": "52b16c67864c8bc5b56a15718c320727e90ffcf7bc4d05c9d3cb28c99979666e",
"md5": "714a3f4676c9768c0d4ac77712284df9",
"sha256": "69945a34437e46164c87d3a96fc5b6eb022131ab8013a4a0b2ea4c0d54b73f24"
},
"downloads": -1,
"filename": "denspa-0.0.3.tar.gz",
"has_sig": false,
"md5_digest": "714a3f4676c9768c0d4ac77712284df9",
"packagetype": "sdist",
"python_version": "source",
"requires_python": ">=3.10",
"size": 8438,
"upload_time": "2025-03-26T16:41:47",
"upload_time_iso_8601": "2025-03-26T16:41:47.649581Z",
"url": "https://files.pythonhosted.org/packages/52/b1/6c67864c8bc5b56a15718c320727e90ffcf7bc4d05c9d3cb28c99979666e/denspa-0.0.3.tar.gz",
"yanked": false,
"yanked_reason": null
}
],
"upload_time": "2025-03-26 16:41:47",
"github": true,
"gitlab": false,
"bitbucket": false,
"codeberg": false,
"github_user": "Kankeu",
"github_project": "denspa",
"travis_ci": false,
"coveralls": false,
"github_actions": false,
"requirements": [
{
"name": "langchain",
"specs": [
[
">=",
"0.3.16"
]
]
},
{
"name": "faiss-cpu",
"specs": [
[
">=",
"1.10.0"
]
]
},
{
"name": "nltk",
"specs": [
[
">=",
"3.9.1"
]
]
},
{
"name": "scikit-learn",
"specs": [
[
">=",
"1.6.1"
]
]
},
{
"name": "numpy",
"specs": [
[
"<=",
"1.26.4"
]
]
},
{
"name": "scipy",
"specs": [
[
">=",
"1.15.1"
]
]
},
{
"name": "spacy",
"specs": [
[
">=",
"3.5.2"
]
]
}
],
"lcname": "denspa"
}
JSON
