MCP 服务器

Semantic Search MCP Server

Provides hybrid semantic and keyword code search for Claude Code using BM25 and vector retrieval. It enables indexing and searching local codebases with language-aware chunking and local embeddings.

README

codesight

AI-powered document search engine — hybrid BM25 + vector + RRF retrieval with Claude answer synthesis.

Quick Start

# Install
pip install -e ".[dev]"

# Index a folder of documents
python -m codesight index /path/to/documents

# Search
python -m codesight search "payment terms" /path/to/documents

# Ask a question (requires ANTHROPIC_API_KEY)
python -m codesight ask "What are the payment terms?" /path/to/documents

# Launch the web chat UI
pip install -e ".[demo]"
python -m codesight demo

Python API

from codesight import CodeSight

engine = CodeSight("/path/to/documents")
engine.index()                                     # Index all files
results = engine.search("payment terms")           # Hybrid search
answer = engine.ask("What are the payment terms?") # Search + Claude answer
status = engine.status()                           # Index freshness check

Supported Formats

Format	Extension	Parser
PDF	`.pdf`	pymupdf
Word	`.docx`	python-docx
PowerPoint	`.pptx`	python-pptx
Code	`.py`, `.js`, `.ts`, `.go`, `.rs`, etc.	Built-in (10 languages)
Text	`.md`, `.txt`, `.csv`	Built-in

Architecture

Document Parsing: PDF, DOCX, PPTX text extraction with page/section metadata
Chunking: Language-aware regex splitting (code) + paragraph-aware splitting (documents)
Embeddings: all-MiniLM-L6-v2 via sentence-transformers (local, no API key)
Vector Store: LanceDB (serverless, file-based)
Keyword Search: SQLite FTS5 sidecar
Retrieval: Hybrid BM25 + vector with RRF merge
Answer Synthesis: Claude API generates answers with source citations

See ARCHITECTURE.md for the full system tour.

Configuration

Variable	Default	Description
`ANTHROPIC_API_KEY`	—	Required for `ask()` / Claude answer synthesis
`CODESIGHT_DATA_DIR`	`~/.codesight/data`	Where indexes are stored
`CODESIGHT_EMBEDDING_MODEL`	`all-MiniLM-L6-v2`	Embedding model
`CODESIGHT_LLM_MODEL`	`claude-sonnet-4-20250514`	Claude model for answers
`CODESIGHT_STALE_MINUTES`	`60`	Index freshness threshold
`LOG_LEVEL`	`INFO`	Logging verbosity

See .env.example for all options.

Workflow: Explore → Plan → Execute → Review

Opus in VS Code plans and launches autonomous CLI agents in the background — the user never leaves the conversation. Agents run via env -u CLAUDECODE claude --dangerously-skip-permissions --model [model] -p '...' with output redirected to files. Multiple cycles ensure quality: Sonnet implements, Opus reviews. See .claude/rules/workflow.md for full details.