# veritasgraph [Health: Active]

**Category:** 🔎 Search & Data Extraction  
**Repository:** https://github.com/bibinprathap/VeritasGraph  
**GitHub Stars:** 325  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/veritasgraph

## Description
Zero-trust, air-gapped Enterprise GraphRAG MCP server with offline, citation-grounded answers.

## Tools
Capabilities this server exposes over MCP:

- **veritasgraph_ingest_document** — Ingest a document into the VeritasGraph knowledge graph. Chunks the text, extracts entities and relationships with a local model, and records the source chunk behind every node/edge for verifiable attribution.
- **veritasgraph_query** — Ask a question and get a graph-grounded, multi-hop answer with verifiable [doc#chunk] citations and the reasoning path used.
- **veritasgraph_search_entities** — Retrieve the subgraph most relevant to a query (entities, relationships, seeds) without invoking the LLM. Fast graph lookup.
- **veritasgraph_get_graph** — Return the full knowledge graph: all nodes, edges, and stats.
- **veritasgraph_clear_graph** — Clear the entire knowledge graph (destructive; removes all data).

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `uvx` (confidence: high):

```json
"mcpServers": {
  "veritasgraph": {
    "command": "uvx",
    "args": ["veritasgraph-mcp"]
  }
}
```

## Documentation

## What veritasgraph does

veritasgraph provides a GraphRAG workflow exposed through MCP for agents that need structured retrieval and evidence-backed answers. Instead of relying only on vector similarity, it combines tree-search organization with knowledge-graph relationships. The project emphasizes provenance recorded during ingestion, reproducible graph traversal, and answers that can abstain when the available evidence is insufficient.

The server is intended for enterprise and governed-agent scenarios where source attribution and local deployment are important. Its examples include medical and pharmaceutical workflows, fact-checking, agent orchestration, and construction-drawing analysis. The construction example turns floorplan PDFs into structured rooms, dimensions, openings, fixtures, and schedules, then builds a graph that supports quantity-takeoff queries with source-sheet provenance.

## How it works

Queries use graph relationships and tree structure to identify connected evidence. The retrieval trace is deterministic, while language-model generation remains separate from the evidence path.

When required graph relationships or supporting evidence are unavailable, the workflow can return an insufficient-evidence result instead of filling the gap with a guess. The MedTrust example extends this approach with mandatory document-and-chunk citations, authorization checks, and a tamper-evident audit log.

## Setup and configuration

Install the package with pip and start the lightweight demonstration with:

```bash
pip install veritasgraph
veritasgraph demo --mode=lite
```

Lite mode uses cloud APIs and requires an OpenAI or Anthropic API key; the quick-start example shows `OPENAI_API_KEY`. Local mode uses Ollama and can run offline, with the example specifying the `llama3.2` model:

```bash
veritasgraph demo --mode=local --model=llama3.2
```

The repository also shows `python -m veritasgraph_mcp` for running from the repository root, with local Ollama required for ingestion and querying. Full mode uses Docker and Neo4j:

```bash
veritasgraph start --mode=full
```

The project targets Python 3.10 or newer. Local use is described as requiring Ollama and 8 GB of RAM.

## Tools and capabilities

- Ingest and query graph-backed knowledge using local Ollama-based workflows.
- Run a cloud-backed lite mode with OpenAI or Anthropic APIs.
- Run an offline mode for privacy-sensitive deployments.
- Use graph traversal and tree-search structure for multi-step retrieval.
- Preserve provenance for citation-grounded responses.
- Abstain when the graph lacks sufficient evidence.
- Support a full deployment using Docker and Neo4j.
- Process floorplan PDFs into graph data and quantity-takeoff outputs in the featured example.

## Limitations and notes

The available modes have different dependencies. Lite mode needs a cloud-provider API key, local mode needs Ollama and suitable local resources, and full mode needs Docker and Neo4j. The repository presents the MedTrust Agent as an example implementation of the governed-agent architecture, rather than as a general description of every deployment.

The server is released under the MIT license. Python 3.10 or later is required for the package.

_Full upstream README: https://allmcps.com/mcp/veritasgraph/readme_