GraphRAG

Entity-centric Retrieval-Augmented Generation for Crypto Whitepapers
Local-first • Private • FastMCP-ready

---

1️⃣ ✨ Overview

GraphRAG MCP is a modular, local-first system that turns crypto whitepapers into an entity-centric Knowledge Graph and a vector-searchable corpus, then answers questions with RAG + optional KG enrichment + LLM synthesis — all via standardized FastMCP tools.

Why this project?

• 🛡️ Privacy by default: runs entirely on your machine (Ollama, Chroma, GraphDB).
• ⚡ Fast & focused: entity-filtered retrieval narrows context to the right tokens/protocols.
• 🧩 Composable: exposes rag.* and kg.* tools so an MCP Coordinator or Streamlit app can orchestrate multi-tool workflows.
• 🧠 Explainable answers: returns citations with doc/chunk/entity IDs for every response.

---

🔁 Typical usage

1. Ingest and label whitepapers → build embeddings and insert entities.
2. Ask questions via rag.qa (semantic + entity-filtered retrieval), optionally enrich with KG labels/aliases.
3. Get concise LLM answers with inline citations to source chunks.

---

2️⃣ Features

🧩 Knowledge Graph (KG)

• Entity-only architecture using RDF/OWL ontologies (mcp-core.ttl, mcp-crypto.ttl).
• Built on Ontotext GraphDB 11+ with SHACL validation and SPARQL/GraphQL endpoints.
• Stores canonical entities such as tokens, protocols, components, and organizations.
• Enables KG enrichment for RAG answers via aliases, labels, and relationships.

🔍 Vector Retrieval (RAG)

• ChromaDB acts as the persistent vector store for chunk embeddings.
• Embeddings generated using Ollama’s nomic-embed-text model.
• Supports semantic and entity-filtered retrieval modes for accurate context fetching.
• Each chunk contains structured metadata: doc_id, chunk_id, entity_ids, section_type, and page.

🧠 Local LLM Inference

• Uses Ollama for fully local inference — no external API keys required.
• Compatible with models like llama3.1:latest, qwen2.5:14b-instruct, or mistral.
• Performs labeling, summarization, and final QA synthesis.
• Includes deterministic mock mode for offline testing and CI.

⚙️ FastMCP Servers

• Two modular servers expose tools via FastMCP 2.x:
  • rag → rag.search, rag.embed_and_index, rag.reindex, rag.delete, rag.health, rag.qa
  • kg → sparql_query, sparql_update, push_labels, validate_labels, list_documents, kg.health
• Both run locally via stdio and are MCP-Coordinator compatible.

🔒 Privacy & Portability

• 100% offline operation — suitable for air-gapped or research environments.
• Reproducible local stack (GraphDB + Chroma + Ollama + FastMCP).
• Works seamlessly on Windows 11, macOS, or Linux.

🚀 Integration Ready

• Plug-and-play with MCP Coordinators or Streamlit apps for end-user Q&A.
• Can interoperate with other MCPs such as:
  • Brave API MCP (web search)
  • MongoDB MCP (strategy data)
  • Telegram MCP (messaging)
  • Gmail MCP (email retrieval)
• Returns clean JSON outputs for easy chaining into agentic workflows.

---

3️⃣ 🏗️ Architecture

The GraphRAG MCP architecture combines Knowledge Graph reasoning, Vector-based retrieval, and Local LLM synthesis — all under the MCP interoperability standard.
It’s designed for clarity, privacy, and modular scalability.

---

🧭 High-Level Overview

Layer	Technology	Purpose	Example Components
🗂 Ingestion Layer	Python + LangChain	Reads PDFs, splits into semantic chunks, labels with LLMs	pdf_reader.py, semantic_splitter.py, llm_chunk_tagger.py
🧩 Knowledge Graph Layer (KG)	GraphDB (Ontotext) + RDFLib	Stores canonical entities (tokens, protocols, organizations)	graphdb_sink.py, namespaces.py, SHACL shapes
💾 Vector Retrieval Layer (RAG)	ChromaDB + Ollama embeddings	Stores text chunks + metadata + embeddings for semantic retrieval	chroma_store.py, .chroma/
⚙️ MCP Layer	FastMCP 2.x	Exposes standardized MCP tools (rag.*, kg.*)	rag_server.py, kg_server.py
🧠 LLM Synthesis Layer	Ollama LLMs (llama3.1, qwen2.5)	Answers questions with retrieved context + KG enrichment	rag.qa, llm_chunk_tagger
💬 User Interface Layer	MCP Coordinator / Streamlit	Connects multiple MCPs for conversational Q&A	Coordinator UI or custom Streamlit dashboard

---

🔹 Data Flow Diagram

            ┌────────────────────────────────────────┐
            │              Whitepapers               │
            │ (PDFs, research papers, documentation) │
            └───────────────────┬────────────────────┘
                                │
                                ▼
            ┌─────────────────────────────────────────────┐
            │      📄 Ingestion & Labeling                │
            │  pdf_reader → semantic_splitter →           │
            │  llm_chunk_tagger → postprocess             │
            └───────────────────┬─────────────────────────┘
                                 │
                        ┌────────┴────────┐
                        │                 │
                        ▼                 ▼
            ┌────────────────┐    ┌──────────────────────┐
            │ 🧠 GraphDB KG  │    │ 💾 Chroma RAG      │
            │ Entities & IRIs │   │ Chunks + Embeddings │
            └────────┬────────┘   └────────┬────────────┘
                     │                     │
                     ▼                     ▼
               ┌───────────────┐      ┌───────────────┐
               │ ⚙️ kg_server  │     │ ⚙️ rag_server │
               │ (FastMCP)     │      │ (FastMCP)     │
               └────────┬──────┘      └──────┬────────┘
                        │                    │
                        └────────┬───────────┘
                                 ▼
                  ┌────────────────────────────────┐
                  │ 💬 MCP Coordinator / Streamlit │
                  │  User-facing Q&A Interface     │
                  └────────────────────────────────┘

---

🧠 How It Works (Step-by-Step)

Step	Description	Input	Output
1️⃣	PDF Parsing	Whitepaper PDF	Raw text pages
2️⃣	Semantic Splitting	Raw text	Meaningful chunks (by section/topic)
3️⃣	LLM Labeling	Chunk text	Entities, relations, and section labels
4️⃣	Postprocessing	Labeled chunks	Cleaned JSONL with canonical entity IRIs
5️⃣	Indexing	JSONL labels	Chroma embeddings + KG triples
6️⃣	Retrieval (rag.search)	Query text / entities	Relevant chunks
7️⃣	Enrichment (optional)	Retrieved entities	KG aliases, definitions
8️⃣	Answer Synthesis (rag.qa)	Question + context	Concise answer with citations

---

🌐 Data Modalities

Data Type	Storage	Example
🧱 Entity	GraphDB	<https://kg.mcp.ai/id/token/bitcoin> → rdf:type crypto:Token
📜 Chunk	Chroma	“Bitcoin is a peer-to-peer electronic cash system…”
🧩 Embedding	Chroma / Ollama	768-dim nomic-embed-text vector
🧮 Provenance	Metadata	doc_id, chunk_id, page, entity_ids[]
💬 Answer	MCP JSON	{ "answer": "...", "citations": [...] }

---

🧱 Core MCP Tools

Server	Tool	Description
🧩 RAG	rag.search	Semantic search over chunks
	rag.embed_and_index	Add new labeled chunks to index
	rag.reindex	Rebuild from outputs directory
	rag.delete	Delete by IDs or filters
	rag.qa	Question answering with LLM synthesis
	rag.health	Diagnostics and store info
🧠 KG	sparql_query / sparql_update	Execute SPARQL against GraphDB
	push_labels / validate_labels	Add or validate KG entries
	list_documents, get_chunk	Retrieve document metadata
	kg.health	Check GraphDB repository status

---

4️⃣ ⚙️ Installation & Setup

Set up your local GraphRAG MCP environment in just a few steps!
This stack runs fully offline and integrates seamlessly with Ollama, GraphDB, and Chroma.

---

🧾 Prerequisites

Requirement	Description	Example
🐍 Python	Version 3.11+ recommended	python --version → Python 3.11.8
🧠 Ollama	Local LLM runtime (for inference + embeddings)	ollama pull llama3.1:latest
🧩 GraphDB Desktop 11+	Local Knowledge Graph database	runs at http://localhost:7200
💾 ChromaDB	Vector store for embeddings	auto-initialized under .chroma/
🧰 FastMCP	Multi-Component Platform runtime (2.x)	installed via pip

---

🧱 Folder Layout (simplified)

Folder	Purpose	Example Contents
src/	Core codebase	pipeline.py, mcp/, kg/, rag/
outputs/run_simple/	Generated outputs	labeled chunks, reports, embeddings
.chroma/	Chroma persistent vector store	chroma.sqlite3, index/
.env	Environment configuration	Ollama, GraphDB, Chroma settings
tests/	Offline unit tests	test_rag_qa.py, test_kg_server.py

---

🧰 Step-by-Step Setup

🪄 1️⃣ Clone & Create Virtual Environment

git clone https://github.com/Swissbit92/GraphDB_Desktop.git

⚡ 2️⃣ Activate Environment

OS	Command
🪟 Windows (PowerShell)	.venv\Scripts\activate
🐧 Linux / macOS	source .venv/bin/activate

📦 3️⃣ Install Dependencies

pip install -r requirements.txt

⚙️ 4️⃣ Verify Installation

python -m src.mcp.rag_server --list-tools
python -m src.mcp.kg_server --list-tools

✅ You should see tools like rag.qa, rag.search, and kg.health.

---

🧠 Optional: Preload Ollama Models

Model	Purpose	Pull Command
🦙 llama3.1:latest	Default reasoning + summarization model	ollama pull llama3.1:latest
🧩 nomic-embed-text	Embedding model for RAG vectorization	ollama pull nomic-embed-text
🤖 qwen2.5:14b-instruct	Larger model for complex QA tasks	ollama pull qwen2.5:14b-instruct

---

🔍 Quick Sanity Check

Run a quick health diagnostic to ensure everything is configured correctly:

pytest -q
python -m src.mcp.rag_server --run-tool rag.health
python -m src.mcp.kg_server --run-tool kg.health

If both return ✅ OK, you’re ready to run the pipeline and start querying your Knowledge Graph + RAG system!

---

5️⃣ 🧪 How to Use & Test

📥 Ingest Whitepapers & Build the Index

# Place your PDFs under .\whitepapers\ then run:
python -m src.pipeline --input ".\whitepapers\*.pdf"

✅ Outputs:

• Labeled JSONL → outputs\run_simple\labels\
• Chroma index → .chroma\
• (If enabled) Entities pushed to GraphDB repository mcp_kg

---

🖧 Start the MCP Servers (RAG + KG)

# Terminal A
python -m src.mcp.rag_server

# Terminal B
python -m src.mcp.kg_server

💡 Tip: In another PowerShell window, confirm the tools are available:

python -m src.mcp.rag_server --list-tools
python -m src.mcp.kg_server --list-tools

---

🔎 Quick Retrieval Check (RAG)

# Example: semantic search for "peer-to-peer electronic cash"
python -m src.mcp.rag_server --run-tool rag.search --input '{ "text": "peer-to-peer electronic cash", "k": 3 }'

You should see matching chunks with doc_id, chunk_id, and distances.

---

❓ Ask Questions with Citations (rag.qa)

# Fully offline (deterministic mock answer)
python -m src.mcp.rag_server --run-tool rag.qa --input '{ "question": "What problem does Bitcoin aim to solve?", "k": 5, "kg_enrich": true, "use_mock_llm": true }'

➡️ Returns:

• answer: concise response (mock or LLM)
• citations: [ {doc_id, chunk_id, entity_ids, text} ]
• took_ms, model_used

Switch to real LLM synthesis by omitting use_mock_llm (requires Ollama running).

---

🧠 Optional: Entity-Filtered QA

python -m src.mcp.rag_server --run-tool rag.qa --input '{ "question": "How does proof-of-work secure the network?", "entity_ids": ["https://kg.mcp.ai/id/token/bitcoin"], "k": 5, "kg_enrich": true, "use_mock_llm": true }'

This restricts retrieval to chunks tagged with the specified KG entity(ies).

---

🧪 Run the Test Suite

pytest -q

Key tests (all offline):

• tests\test_rag_qa.py: verifies retrieval normalization and mock LLM mode
• tests\test_kg_server.py: checks KG connectivity (skips if GraphDB not running)

---

🩺 Health Checks

python -m src.mcp.rag_server --run-tool rag.health
python -m src.mcp.kg_server --run-tool kg.health

Expect collection info, document counts, and OK status.

---

🧩 MCP Coordinator / UI Hookup (Optional)

Ensure your mcp.json references the running servers:

{
  "mcpServers": {
    "rag": { "command": "python", "args": ["-m", "src.mcp.rag_server"] },
    "kg":  { "command": "python", "args": ["-m", "src.mcp.kg_server"] }
  }
}

Then connect via your MCP Coordinator or Streamlit app to interactively call rag.qa and kg.* tools.

---

🙏 Closing Words

GraphRAG MCP is part of the broader Eeva AI ecosystem — an open, modular framework for intelligent crypto research and strategy generation.
This project wouldn’t exist without the incredible open-source community that continues to push the boundaries of local AI and knowledge engineering.

If you find this useful:

• ⭐ Star the repository to support ongoing development
• 🧩 Contribute improvements or new MCP modules
• 🧠 Explore integrations with other MCPs (Brave API, MongoDB, Telegram, etc.)
• 💬 Share feedback — every suggestion helps make the system smarter, faster, and more reliable

---

“Knowledge is only powerful when it’s connected.”
— Eeva AI Research

Thank you for being part of the open-source journey. 🚀

---