smart-search
A local-first MCP server that enables semantic search over PDF and DOCX documents using structure-aware parsing and vector storage. It allows users to query their local knowledge base through Claude Code without cloud dependencies or GPU requirements.
README
smart-search
A local-first MCP server that makes PDF and DOCX files searchable from Claude Code via semantic search. Documents are extracted with Docling, chunked using structure-aware hierarchical chunking, embedded with nomic-embed-text-v1.5 (ONNX), and stored in LanceDB. Everything runs on CPU with no cloud dependencies and no GPU required.
Version: 0.1.0 (Foundation)
MCP Tools
The server exposes two tools to Claude Code.
knowledge_search
Search the knowledge base for document chunks matching a natural language query.
| Parameter | Type | Required | Default | Description |
|---|---|---|---|---|
query |
string | Yes | - | Natural language search query |
limit |
integer | No | 10 |
Maximum number of chunks to return |
mode |
string | No | "hybrid" |
Search mode: semantic, keyword, or hybrid |
doc_types |
list of strings | No | null |
Filter by file type, e.g. ["pdf"] or ["pdf", "docx"] |
Returns a formatted block with source file name, page number, section heading path, chunk text, and relevance score for each result.
Note (v0.1): All three mode values currently execute semantic search. Keyword and hybrid modes using SQLite FTS5 + Reciprocal Rank Fusion are planned for v0.3.
knowledge_stats
Returns counts and metadata about the indexed knowledge base: document count, chunk count, index size in MB, last indexed timestamp, and file formats present.
No parameters.
Prerequisites
- Python 3.11 or later
uv(recommended) orpip
The embedding model (nomic-ai/nomic-embed-text-v1.5) is downloaded from Hugging Face on first run and cached locally. No GPU or internet connection is required after that initial download.
Installation
Clone the repository and install in editable mode with development dependencies:
git clone <repository-url>
cd smart-search
uv pip install -e ".[dev]"
To install without development dependencies:
uv pip install -e .
Indexing Documents
Documents must be indexed before they can be searched. Use DocumentIndexer directly from Python, or integrate it into a CLI or script.
Index a single file:
from smart_search.config import get_config
from smart_search.chunker import DocumentChunker
from smart_search.embedder import Embedder
from smart_search.store import ChunkStore
from smart_search.indexer import DocumentIndexer
config = get_config()
store = ChunkStore(config)
store.initialize()
indexer = DocumentIndexer(
config=config,
chunker=DocumentChunker(config),
embedder=Embedder(config),
store=store,
)
result = indexer.index_file("/path/to/document.pdf")
print(result.status, result.chunk_count)
Index a folder:
result = indexer.index_folder("/path/to/documents", recursive=True)
print(f"Indexed: {result.indexed}, Skipped: {result.skipped}, Failed: {result.failed}")
Files already indexed at the same content hash are skipped automatically. Pass force=True to re-index regardless.
MCP Server Setup
Configure Claude Code
Copy .mcp.json.example to .mcp.json and update the paths to match your environment:
{
"mcpServers": {
"smart-search": {
"command": "/path/to/your/venv/bin/python",
"args": ["-m", "smart_search.server"],
"cwd": "/path/to/smart-search"
}
}
}
On Windows, use the full path to python.exe inside your virtual environment's Scripts directory.
Run the server manually
python -m smart_search.server
The server communicates over stdio (MCP standard transport).
Configuration
All settings can be overridden with environment variables prefixed SMART_SEARCH_. The defaults are suitable for local use out of the box.
| Environment Variable | Default | Description |
|---|---|---|
SMART_SEARCH_EMBEDDING_MODEL |
nomic-ai/nomic-embed-text-v1.5 |
Hugging Face model identifier |
SMART_SEARCH_EMBEDDING_DIMENSIONS |
768 |
Output vector dimensions |
SMART_SEARCH_EMBEDDING_BACKEND |
onnx |
Backend: onnx (default) or pytorch |
SMART_SEARCH_CHUNK_MAX_TOKENS |
512 |
Maximum tokens per chunk |
SMART_SEARCH_LANCEDB_PATH |
./data/vectors |
Directory for LanceDB vector store |
SMART_SEARCH_SQLITE_PATH |
./data/metadata.db |
Path to SQLite metadata database |
SMART_SEARCH_LANCEDB_TABLE_NAME |
chunks |
LanceDB table name |
SMART_SEARCH_SEARCH_DEFAULT_LIMIT |
10 |
Default result count for knowledge_search |
SMART_SEARCH_SEARCH_DEFAULT_MODE |
hybrid |
Default search mode for knowledge_search |
SMART_SEARCH_NOMIC_DOCUMENT_PREFIX |
search_document: |
Task prefix applied to document text at index time |
SMART_SEARCH_NOMIC_QUERY_PREFIX |
search_query: |
Task prefix applied to queries at search time |
Paths are resolved to absolute paths at startup, so relative values are interpreted relative to the working directory where the server process starts.
Running Tests
The test suite uses pytest. Tests are split into fast unit tests and slow integration tests that load ML models or process real files.
Run fast tests only (default):
pytest
Run all tests including slow ones:
pytest --override-ini="addopts="
Run with coverage:
pytest --cov=smart_search --cov-report=term-missing
Slow tests are marked with @pytest.mark.slow. The default pytest invocation excludes them so the suite completes quickly without loading ML models.
Project Structure
src/smart_search/
server.py - FastMCP entry point; registers knowledge_search and knowledge_stats
indexer.py - Document ingestion pipeline (chunk, embed, store, dedup)
chunker.py - Docling DocumentConverter and HierarchicalChunker wrapper
embedder.py - nomic-embed-text-v1.5 ONNX embedding generation
store.py - LanceDB vector store and SQLite metadata store
search.py - Semantic search with Smart Context result formatting
models.py - Pydantic models: Chunk, SearchResult, IndexStats
config.py - Settings with SMART_SEARCH_ environment variable overrides
tests/
test_models.py - Chunk, SearchResult, IndexStats validation
test_config.py - Environment variable override and path resolution
test_chunker.py - DocumentChunker (slow: requires Docling)
test_embedder.py - Embedder (slow: loads ONNX model)
test_store.py - ChunkStore LanceDB and SQLite operations
test_indexer.py - DocumentIndexer pipeline integration
test_search.py - SearchEngine formatting and filtering
test_server.py - FastMCP tool registration and dispatch
Tech Stack
| Component | Library / Model |
|---|---|
| MCP server | FastMCP |
| Document parsing | Docling (DocumentConverter, HierarchicalChunker) |
| Embeddings | nomic-ai/nomic-embed-text-v1.5 via sentence-transformers + ONNX |
| Vector store | LanceDB |
| Metadata store | SQLite |
| Config | pydantic-settings |
| Build | Hatchling |
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。