TigerGraph GraphRAG v2.0.4

GraphRAG 2.0.4 enhances the Agentic planner for smaller models and adds TigerGraph's query checker to query generation.

What's New

Agentic planner enhancements for smaller models — each plan is now checked for search or answer steps; a plan without steps is made again, and if it still has none, a default search plan is used. Repeated text in a plan is cut. The planner now reviews the context it is given — the conversation and the graph's schema — before it plans: it decides whether that context already answers the question, which parts are still missing, and whether an earlier answer that found nothing, was incomplete, or is out of date must be searched again. Questions about the knowledge base itself, such as "what data do we have", are answered from the schema. The planner now receives the graph's schema and the most recent conversation turns. Tested with GPT-5.4 mini and Gemini 3.5 Flash.
Query generation enhancements with TigerGraph's query checker — each generated query is now checked with TigerGraph's query checker before it runs. A query that fails the check is not run; its errors go back to the query generator for the next attempt, up to 10 attempts by default (cypher_max_attempts). The query generator now follows TigerGraph's rules for dates and missing values, and queries that change data are rejected.

Unanswered questions — when no answer, or only a partial answer, is found, the response says so and suggests asking again with more specific details.
Trace Logs — query steps now show the query that ran, the number of rows returned, and each failed attempt with its error; plan steps show the context review; and a failed answer ends with a step that shows the error.
Service logs — routine requests, such as loading conversations or opening a trace, now write one line to the service log instead of four.

External MCP servers not loading — fixed servers configured on the MCP Servers page failing to load, which left their tools unavailable to the agent.
Empty query results treated as data — fixed a query that returned no rows being treated as found data, which stopped the fallback to document search and left the question unanswered.
Failing step before each query — fixed every query first running a step that always failed, which added a model call and delay before the query was generated.
Trace Logs window overflowing the screen — fixed a long line, such as a URL or a query, widening the window past the screen and hiding the Close button.

Gemini 3 models — Agentic planning with Gemini 3 models uses temperature 1.0 when a lower value is configured, as Google recommends for these models. Answers and knowledge-graph building keep the configured temperature.
Web server access log — the per-request access log is no longer written; each request is still recorded once in the service log.