TrustGraph streamlines the delivery and management of complex AI environments, acting as a comprehensive provisioning platform for your containerized AI tools, pipelines, and integrations.
Deploying state-of-the-art AI requires managing a complex web of models, frameworks, data pipelines, and monitoring tools. TrustGraph simplifies this by providing a unified, open-source solution to provision complete, trusted AI environments anywhere you need them – from cloud instances and on-premises servers to edge devices.
Table of Contents
- Unified Provisioning: Define and deploy complete AI environments, including models, dependencies, and tooling, as a single, manageable unit. Stop managing piecemeal installations.
- No-code TrustRAG Pipelines: Deploy full end-to-end RAG pipelines using unique TrustGraph algorithms leveraging both Knowledge graphs and VectorDBs.
- Environment-Agnostic Deployment: Provision consistently across diverse infrastructures (Cloud, On-Prem, Edge, Dev environments). Build once, provision anywhere.
- Trusted & Secure Delivery: Focuses on providing a secure supply chain for AI components.
- Simplified Operations: Radically reduce the complexity and time required to stand up and manage sophisticated AI stacks. Get operational faster.
- Open Source & Extensible: Built with transparency and community collaboration in mind. Easily inspect, modify, and extend the platform to meet your specific provisioning needs.
- Component Flexibility: Avoid component lock-in. TrustGraph integrates multiple options for all system components.
See the API Developer's Guide for more information.
For users, TrustGraph has the following interfaces:
The TrustGraph CLI installs the commands for interacting with TrustGraph while running along with the Python SDK. The Configuration Builder enables customization of TrustGraph deployments prior to launching. The REST API can be accessed through port 8088 of the TrustGraph host machine with JSON request and response bodies.
pip3 install trustgraph-cli==0.21.17
Note
The TrustGraph CLI version must match the desired TrustGraph release version.
TrustGraph is endlessly customizable by editing the YAML launch files. The Configuration Builder provides a quick and intuitive tool for building a custom configuration that deploys with Docker, Podman, Minikube, AWS, Azure, Google Cloud, or Scaleway. There is a Configuration Builder for the both the lastest and stable TrustGraph releases.
The Configuration Builder has 4 important sections:
- Component Selection ✅: Choose from the available deployment platforms, LLMs, graph store, VectorDB, chunking algorithm, chunking parameters, and LLM parameters
- Customization 🧰: Customize the prompts for the LLM System, Data Extraction Agents, and Agent Flow
- Test Suite 🕵️: Add the Test Suite to the configuration available on port
8888 - Finish Deployment 🚀: Download the launch
YAMLfiles with deployment instructions
The Configuration Builder will generate the YAML files in deploy.zip. Once deploy.zip has been downloaded and unzipped, launching TrustGraph is as simple as navigating to the deploy directory and running:
docker compose up -d
Tip
Docker is the recommended container orchestration platform for first getting started with TrustGraph.
When finished, shutting down TrustGraph is as simple as:
docker compose down -v
The -v flag will destroy all data on shut down. To restart the system, it's necessary to keep the volumes. To keep the volumes, shut down without the -v flag:
docker compose down
With the volumes preserved, restarting the system is as simple as:
docker compose up -d
All data previously in TrustGraph will be saved and usable on restart.
If added to the build in the Configuration Builder, the Test Suite will be available at port 8888. The Test Suite has the following capabilities:
- Graph RAG Chat 💬: Graph RAG queries in a chat interface
- Vector Search 🔎: Semantic similarity search with cosine similarity scores
- Semantic Relationships 🕵️: See semantic relationships in a list structure
- Graph Visualizer 🌐: Visualize semantic relationships in 3D
- Data Loader 📂: Directly load
.pdf,.txt, or.mdinto the system with document metadata
TrustGraph is fully containerized and is launched with a YAML configuration file. Unzipping the deploy.zip will add the deploy directory with the following subdirectories:
docker-composeminikube-k8sgcp-k8s
Note
As more integrations have been added, the number of possible combinations of configurations has become quite large. It is recommended to use the Configuration Builder to build your deployment configuration. Each directory contains YAML configuration files for the default component selections.
Docker:
docker compose -f <launch-file.yaml> up -d
Kubernetes:
kubectl apply -f <launch-file.yaml>
TrustGraph is designed to be modular to support as many LLMs and environments as possible. A natural fit for a modular architecture is to decompose functions into a set of modules connected through a pub/sub backbone. Apache Pulsar serves as this pub/sub backbone. Pulsar acts as the data broker managing data processing queues connected to procesing modules.
TrustGraph incorporates TrustRAG, an advanced RAG approach that leverages automatically constructed Knowledge Graphs to provide richer and more accurate context to LLMs. Instead of relying solely on unstructured text chunks, TrustRAG understands and utilizes the relationships between pieces of information.
How TrustRAG Works:
-
Automated Knowledge Graph Construction:
- TrustGraph processes source data to automatically extract key entities, topics, and the relationships connecting them.
- It then maps these extracted semantic relationships and concepts to high-dimensional vector embeddings, capturing the nuanced meaning beyond simple keyword matching.
-
Hybrid Retrieval Process:
- When a query is received, TrustRAG first performs a cosine similarity search on the vector embeddings to identify potentially relevant concepts and relationships within the knowledge graph.
- This initial vector search pinpoints relevant entry points within the structured Knowledge Graph.
-
Context Generation via Subgraph Traversal:
- Based on the ranked results from the similarity search, TrustRAG dynamically generates relevant subgraphs.
- It starts from the identified entry points and traverses the connections within the Knowledge Graph. Users can configure the number of 'hops' (relationship traversals) to expand the contextual window, gathering interconnected information.
- This structured subgraph, containing entities and their relationships, forms a highly relevant and context-aware input prompt for the LLM that is endlessly configurable with options for the number of entities, relationships, and overall subgraph size.
One of the biggest challenges currently facing RAG architectures is the ability to quickly reuse and integrate knowledge sets. TrustGraph solves this problem by storing the results of the document ingestion process in reusable Knowledge Cores. Being able to store and reuse the Knowledge Cores means the process has to be run only once for a set of documents. These reusable Knowledge Cores can be loaded back into TrustGraph and used for TrustRAG.
A Knowledge Core has two components:
- Set of Graph Edges
- Set of mapped Vector Embeddings
When a Knowledge Core is loaded into TrustGraph, the corresponding graph edges and vector embeddings are queued and loaded into the chosen graph and vector stores.
As a full-stack platform, TrustGraph provides all the stack layers needed to connect the data layer to the app layer for autonomous operations.
TrustGraph seamlessly integrates API services, data stores, observability, telemetry, and control flow for a unified platform experience.
- LLM Providers: Anthropic, AWS Bedrock, AzureAI, AzureOpenAI, Cohere, Google AI Studio, Google VertexAI, Llamafiles, LM Studio, Mistral, Ollama, and OpenAI
- Vector Databases: Qdrant, Pinecone, and Milvus
- Knowledge Graphs: Memgraph, Neo4j, and FalkorDB
- Data Stores: Apache Cassandra
- Observability: Prometheus and Grafana
- Control Flow: Apache Pulsar
- For control flows, Pulsar accepts the output of a processing module and queues it for input to the next subscribed module.
- For services such as LLMs and embeddings, Pulsar provides a client/server model. A Pulsar queue is used as the input to the service. When processed, the output is then delivered to a separate queue where a client subscriber can request that output.
TrustGraph extracts knowledge documents to an ultra-dense knowledge graph using 3 automonous data extraction agents. These agents focus on individual elements needed to build the knowledge graph. The agents are:
- Topic Extraction Agent
- Entity Extraction Agent
- Relationship Extraction Agent
The agent prompts are built through templates, enabling customized data extraction agents for a specific use case. The data extraction agents are launched automatically with the loader commands.
PDF file:
tg-load-pdf <document.pdf>
Text or Markdown file:
tg-load-text <document.txt>
Once the knowledge graph and embeddings have been built or a cognitive core has been loaded, RAG queries are launched with a single line:
tg-invoke-graph-rag -q "What are the top 3 takeaways from the document?"
Invoking the Agent Flow will use a ReAct style approach the combines Graph RAG and text completion requests to think through a problem solution.
tg-invoke-agent -v -q "Write a blog post on the top 3 takeaways from the document."
Tip
Adding -v to the agent request will return all of the agent manager's thoughts and observations that led to the final response.
Once the platform is running, access the Grafana dashboard at:
http://localhost:3000
Default credentials are:
user: admin
password: admin
The default Grafana dashboard tracks the following:
- LLM Latency
- Error Rate
- Service Request Rates
- Queue Backlogs
- Chunking Histogram
- Error Source by Service
- Rate Limit Events
- CPU usage by Service
- Memory usage by Service
- Models Deployed
- Token Throughput (Tokens/second)
- Cost Throughput (Cost/second)
TrustGraph is licensed under AGPL-3.0.
