convez.ai
MCP Server
shashhwatiwari

iSearch4u

shashhwatiwari/iSearch4u

Ver no GitHub
InicialPluginScore · de 0 a 100

Local(for now) multi-agent research assistant - LangGraph pipeline with semantic retrieval over your documents, exposed as an MCP tool for Claude Desktop

Instalação

$npx convez-ai shashhwatiwari/iSearch4u

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

README

A tradução para o português deste README ainda está na fila.

iSearch4u - Multi-Agent Research Assistant

A local research assistant that ingests your documents, stores them in a vector database, and answers questions using a multi-agent LangGraph pipeline. Callable from Claude Desktop via MCP.


What it does

  1. You drop .md files into a docs/ folder and run the ingestion script
  2. Documents are chunked, embedded, and stored persistently in ChromaDB
  3. When you ask a question, a 4-node agent pipeline runs:
    • Planner - breaks your question into focused sub-queries
    • Retriever - searches ChromaDB for relevant chunks
    • Synthesizer - writes a grounded answer using only the retrieved context
    • Critic - evaluates the answer and loops back to Retriever if quality is poor
  4. The whole pipeline is exposed as an MCP tool, callable from Claude Desktop

Tech stack

ComponentTool
LanguagePython 3.10+
LLMGemini API (google-genai)
Agent orchestrationLangGraph
Vector databaseChromaDB (local, persistent)
Embeddingssentence-transformers (all-MiniLM-L6-v2)
Server interfaceMCP (Model Context Protocol)

Project structure

iSearch4u/
├── agent/
│   ├── graphV1.py        # Synthesizer only
│   ├── graphV2.py        # Retriever + Synthesizer
│   ├── graphV3.py        # Planner + Retriever + Synthesizer
│   └── graphV4.py        # + Critic with conditional loop (current)
├── ingestion/
│   ├── ingest.py         # Chunk, embed, and store documents
│   └── query.py          # Test semantic search directly
├── StarterKit/
│   ├── simple_call.py    # Basic Gemini API call
│   ├── looped_call.py    # Multi-turn chat loop
│   └── tool_agent.py     # Tool-use agent loop from scratch
├── docs/                 # Drop your .md files here
├── mcp_server.py         # MCP server - wraps the pipeline as a tool
├── .env.example          # Environment variable template
└── .gitignore

Setup

1. Clone the repo and install dependencies

pip install mcp chromadb sentence-transformers langgraph google-genai python-dotenv

2. Add your Gemini API key

cp .env.example .env
# then open .env and add your key

3. Add documents

Drop .md files into the docs/ folder.

4. Ingest documents

python ingestion/ingest.py

5. Test retrieval

python ingestion/query.py

6. Run the pipeline directly

python agent/graphV4.py

Claude Desktop setup (MCP)

Add this to your Claude Desktop config at ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "iSearch4u": {
      "command": "/path/to/your/python3",
      "args": ["/absolute/path/to/iSearch4u/mcp_server.py"],
      "cwd": "/absolute/path/to/iSearch4u"
    }
  }
}

Find your Python path with which python3. Restart Claude Desktop after saving.


How the agent graph works

User question
      ↓
  [ Planner ]   - asks LLM to generate 1-3 focused sub-queries
      ↓
  [ Retriever ] - embeds each sub-query, searches ChromaDB, deduplicates chunks
      ↓
  [ Synthesizer ] - answers using only retrieved context
      ↓
  [ Critic ]    - verdict: GOOD → done | BAD → loop back to Retriever (max 3 iterations)
      ↓
   Answer

State flows through every node as a typed dictionary - each node reads what it needs and writes back only what it produced.


Environment variables

VariableDescription
GEMINI_API_KEYYour Gemini API key from Google AI Studio
Ver o original no GitHub

Bom para

Relacionados