TrustGraph
Back to News
Release 2.6

TrustGraph 2.6: Next-Gen Cross-Encoder Reranking, New Domain Explorers, and Advanced Enterprise IAM

July 20, 2026
6 min read

San Francisco, CA — TrustGraph 2.6 is here, marking a massive leap forward in retrieval accuracy, enterprise access control, and UI/UX experience. This release completely overhauls our GraphRAG and Document-RAG scoring mechanisms with a highly efficient FlashRank cross-encoder, introduces a suite of powerful new visualization explorers in the UI, and expands support for global enterprise IAM and multiple LLM providers.

Whether you're an enterprise architect fine-tuning access controls or a developer building advanced RAG applications, TrustGraph 2.6 delivers the performance, flexibility, and precision you need.

Major Features & Backend Upgrades

Cross-Encoder Reranking for GraphRAG & Document-RAG

We've completely replaced the legacy three-prompt LLM edge scoring pipeline (kg-edge-*) with a highly optimized FlashRank cross-encoder reranker service.

  • GraphRAG: Introduces a new hop_and_filter() method for iterative graph traversal with per-hop semantic scoring. Direction-aware reranker text avoids duplicate evaluations, and schema predicates are automatically filtered.
  • Document-RAG: Features a two-limit control system (fetch_limit and doc_limit). It over-fetches a candidate pool from the vector store, reranks with the cross-encoder, and applies optional Maximal Marginal Relevance (MMR) for diversity selection, drastically reducing redundant chunks.
  • Backward Compatible: Reranking acts as a no-op if no reranker role is wired, ensuring a smooth transition.

OpenAI Processor API Variants

A single OpenAI-compatible processor can now cover multiple providers via the --variant flag. Supported profiles include openai, deepseek, qwen, mistral, llama, dashscope, and glm. Each variant intelligently handles provider-specific differences (output token parameters, reasoning toggles, temperature rules). Additionally, a new --thinking flag (off/low/medium/high) has been added for reasoning-capable models.

Portable Workspace Export/Import Bundles

Migrating workspaces between TrustGraph deployments is now seamless with portable .tgx bundles. Using the new tg-export-workspace and tg-import-workspace CLI commands, you can export workspace configurations as JSON, triples as N-Quads, and full document libraries into a streamed, gzipped tar archive. Imports are additive for knowledge and skip-existing for config.

Full RDF Language Tag & Datatype Preservation

TrustGraph now fully preserves RDF literal language tags and XSD datatypes through the entire ingestion and query pipeline. The triple dataclass and wire format have been updated to carry o_datatype and o_language, enabling accurate multilingual dataset handling and FILTER(LANG()) SPARQL queries.

Enterprise IAM & Global User Model

  • Fine-Grained Capabilities: Coarse gateway capabilities have been split into per-operation variants (e.g., graph:read → triples:read, sparql:read, graph-rag:read). Enterprise IAM schema extensions now support group_id, member_type, and grant fields.
  • Global Username Lookup: Fixed the Cassandra user table and JWT claims to match a global user model. iam_users_by_username now drops the workspace from its PRIMARY KEY, meaning logins are global rather than per-workspace. UserRecord.workspace has been renamed to default_workspace throughout the stack.

Guided macOS Installer

Setting up TrustGraph on macOS is easier than ever. The interactive install_trustgraph.sh script detects hardware, recommends an LLM mode (OpenAI vs. Ollama), installs prerequisites via Homebrew, sets up a Python venv, and health-checks the API gateway.

UI Updates (v0.3.12)

The TrustGraph UI receives a massive upgrade in 2.6, shifting to client-side routing with react-router so browser back/forward buttons, deep links, and page refreshes all work flawlessly.

Powerful New Domain Explorers & Visualizations

We've introduced a suite of specialized explorers to help you navigate complex datasets visually:

  • Subway-Map Pathway Finder: Replaced flat per-path cards with a merged, interactive subway-map DAG visualization, revealing hub nodes and chokepoints at a glance.
  • Innovation Intelligence Explorer: A full-featured ecosystem explorer featuring Browse, Pathway Finder (DFS-based pathfinding), and a GTM Advisor for streamed strategy report generation.
  • Risk Explorer: A master-detail viewer showing actors, risks, assets, and events with risk-score bars and navigable relationship links.
  • Game Theory Explorer: Interactive game tree visualization with backward induction, Nash Equilibrium detection, payoff matrix heatmaps, and a sandbox mode with live sliders.
  • Law in Context Explorer: A five-mode legal explorer (overview, institutions, rights, compliance, structure) with EN/LT language toggle and multi-law browsing.

UX Improvements & Demos

  • Added Brand Analytics and Retail demo pages (featuring decision traces and product hover cards).
  • Updated explainability views to match the backend's switch to cross-encoder reranking (tg:concept and tg:score).
  • Workspace recovery from stale sessionStorage, delete/discard buttons on Document Ingest, and a fielded input mode for the prompt test panel.

Bug Fixes & Stability

Backend

  • JWT Signing Key Retry: Fixed an issue where the API gateway would permanently fail JWT auth (returning 401) if IAM wasn't ready on startup; it now lazily retries.
  • Literal Object Indexing: Fixed a Cassandra issue where quads_by_entity skipped literal objects, causing tg-query-graph to return no results.
  • Missing Time Field: Resolved a KeyError crash in the Library API when document records lacked a stored time field.
  • Mux Authorisation: Workspace-scoped operations now correctly resolve workspace from the envelope only, preventing spurious access-denied errors.

UI

  • Agent DAG Hang & Retry Storms: Fixed infinite loops in the Full DAG explainability view caused by back-edges, and switched to single-attempt agent queries with a 180s timeout.
  • Streaming Error Crash: Normalized gateway top-level error objects to prevent React from crashing on streaming receivers.
  • RAG Streaming Timeouts: Increased GraphRAG and DocumentRAG streaming timeouts to 180s to match agent settings.
  • Upload Chunk Size: Reduced upload chunk size from 5MB to 2MB to safely stay within Pulsar’s message size limit after encoding.
  • Triple Writer Reconnect: WebSocket connections now only open when there are triples to flush, halting endless 2s reconnect loops.

Breaking Changes

Please review the following breaking changes before upgrading:

  1. GraphRAG Pipeline: The three kg-edge-* LLM prompt services and the edge_score_limit parameter are removed. Deployments must wire the new reranker service.
  2. RDF Round-trip: Existing ingested data will lose language/datatype metadata. Re-ingestion is required to backfill language tags and datatypes.
  3. JWT Claim Rename: "workspace" is now "default_workspace" in JWT claims, wire protocol, and Cassandra schema. Re-login and schema migration are required.
  4. Fine-Grained Capabilities: Custom policy configs using hardcoded coarse capability strings (graph:read, graph:write, documents:read, documents:write, rows:read) must be updated to the new fine-grained names.

Get Started

TrustGraph 2.6 is available now on GitHub. Check out the updated documentation to learn how to wire up the new cross-encoder reranker, configure LLM variants, or try out the new UI explorers!

Watch TrustGraph 2.6 in Action

For more information:

View all news articles