trustgraph

San Francisco, CA — TrustGraph 2.8 is here. This release delivers a foundational platform upgrade for teams deploying knowledge-driven AI systems at enterprise scale: fully asynchronous messaging infrastructure, hybrid Document RAG retrieval, native LLM structured output, image understanding, comprehensive auditability, and a redesigned extensible UI platform.

TrustGraph 2.8 removes a major scaling constraint in multi-workspace deployments while expanding the ways developers can ingest, retrieve, govern, observe, and present contextual AI applications.

Whether you are operating GraphRAG across dozens of workspaces, building multimodal enterprise search, enforcing reliable JSON outputs from LLMs, or tailoring an investigation interface to your domain, TrustGraph 2.8 provides a faster and more composable foundation.

Major Features & Platform Upgrades

Async Pub/Sub Infrastructure for Enterprise Scale

TrustGraph has completed its migration from a thread-per-consumer messaging model to an asynchronous pub/sub architecture.

The previous model could create hundreds of threads as workspace and flow counts increased. TrustGraph 2.8 replaces that approach with configurable async receive and send pools, dramatically reducing resource pressure while supporting larger deployments.

• Async implementations are now available across Pulsar, RabbitMQ, and Kafka.
• The API gateway and reverse gateway now run on async infrastructure.
• The migration removes approximately 7,600 lines of synchronous pub/sub code.
• Deployments that previously faced thread exhaustion at 40 or more workspaces can now use pooled async consumers and producers instead.

This is a foundational change for organizations operating TrustGraph as a multi-tenant or high-workspace platform.

Hybrid Retrieval for Document RAG

Document RAG now combines semantic vector search with BM25 keyword search through Reciprocal Rank Fusion (RRF).

Vector retrieval is strong at finding conceptually similar material. Keyword retrieval is strong when exact terms, identifiers, error codes, names, and specialist vocabulary matter. Hybrid retrieval combines both signals so that retrieval can better handle real enterprise content.

TrustGraph 2.8 introduces:

• A pluggable KeywordIndexService.
• SQLite FTS5 as the initial keyword-index backend.
• Index scoping per workspace and collection.
• Text sanitisation for safe and reliable BM25 indexing.
• Graceful fallback when the keyword-retrieval path is unavailable.

The result is more resilient document retrieval across both semantic questions and exact-term queries.

LLM-Native Structured Output

TrustGraph 2.8 brings JSON-schema enforcement directly into the LLM generation path.

Schemas can now flow from prompt definitions through the text-completion service to supported providers’ native structured-output APIs. A runtime compatibility check determines whether a prompt and backend can use native schema enforcement.

Structured output support now spans:

• OpenAI, including the vLLM variant.
• Ollama.
• Anthropic Claude.
• Mistral.
• Llamafile and LM Studio.
• Azure OpenAI and Azure serverless.
• Text Generation Inference.
• Vertex AI Gemini and Google AI Studio.
• Amazon Bedrock.

This makes it easier to build dependable agents, extraction pipelines, and user interfaces that consume typed LLM responses rather than parsing unpredictable free-form text.

Image-to-Text Service

TrustGraph now includes a pluggable image description service for multimodal applications.

The first implementation includes an OpenAI vision backend, gateway dispatch, REST and WebSocket clients, IAM capability integration, and a new tg-describe-image command-line tool.

Teams can use the service to turn visual content into text that can be stored, indexed, queried, or incorporated into downstream knowledge and AI workflows.

Docling Document Decoder