Release v2.0.0

TigerGraph GraphRAG v2.0.0

GraphRAG 2.0.0 is a major release built around a new agentic chat engine, structure-aware document chunking, and additive prompt customization.

What's New

• Agentic chat engine — a new default engine that plans and runs its own retrieval instead of following a fixed pipeline. It comes in two styles — Planned (lays out the whole retrieval plan up front, then answers) and Reactive (decides each step from the previous result) — chooses which methods to use, answers greetings and questions about itself directly, and cites the chunks and queries behind each answer. The Classic single-method engine remains available.
• External MCP tools — superusers can register external MCP servers (and install the Python packages they need) so the agent can call their tools during a conversation.
• Structure-aware document chunking — Markdown and HTML are split by structure, keeping each section's heading context in the chunk, rolling small sections into their parent up to the size budget, and keeping tables intact, so retrieval holds together on heading- and table-heavy documents.
• Additive prompt customization — the Customize Prompts page exposes an editable instructions-and-examples layer over fixed rules, including separate, editable retrieval strategies for the Planned and Reactive agents; edits extend behavior without dropping the rules that keep prompts working.

Changed

• Retrieval is more reliable on table-heavy and numeric content — each chunk is embedded together with a compact summary of its topic, section, and key entities, so its vector carries that context explicitly.
• Search results are bounded by relevance — hybrid and community search return the most relevant chunks up to a configurable limit, rather than everything the graph expansion reaches, reducing the context sent to the model. Tunable on the GraphRAG Configuration page.
• Answers stay grounded in document text — the agent always includes a vector search unless a question is confidently a pure structured-data request (an exact count, lookup, relationship, or aggregation).
• A streaming answer can be stopped — while the agent responds, the chat's send button becomes a stop control that ends the current response and re-enables the input.
• Chat and admin UI refinements — a clearer chat engine/style picker, bulk-clearing of older conversations, incremental chat-history loading, and the graph Compatibility Check renamed Migration Assistant.
• More reliable query installation on large graphs — graph queries install through a non-blocking request with status polling, so initialization no longer times out while queries compile.
• Ingestion survives a transient database disconnect — files whose load hits a connection error are retried once the database is reachable again, and any that still fail are named so a re-run reloads only those.

Bug Fixes

• A single oversized chunk no longer drops embeddings for the rest of its batch — over-limit embeddings are retried at progressively shorter lengths, and a vertex that still doesn't fit is skipped on its own instead of aborting the batch.
• Large ingests no longer fail on oversized upsert batches — upserts are sized to the pending work, and progress counts reflect distinct vertices and edges.
• Non-ASCII answers no longer truncate incorrectly on large context — retrieved context is measured in the same form that is sent to the model, so Japanese and other multi-byte content is sized correctly.

Upgrade notes

• Query responses default to answer-only — the query endpoints no longer return supporting sources and trace unless the caller opts in; integrations that read those fields must request them explicitly.
• Existing prompt customizations must be re-saved — with the move to additive customization, pre-existing full-prompt overrides are ignored until re-saved in the new instructions-and-examples form.
• Pick up query fixes on existing graphs — after upgrading, use Knowledge Graph Admin → Migration Assistant to re-create and reinstall shipped queries whose body has drifted, without rebuilding.

Full Changelog

v1.4.2...v2.0.0