Repository navigation
GML-2323 Release GraphRAG 2.0.4 - #72
Merged
Merged
Conversation
- Ask for one short sentence in the plan summary and step reasons, and shorten an overlong value instead of failing the plan - Cap the planner's output tokens, with a new planner_max_output_tokens setting and a higher floor for reasoning models Refs: GML-2324
- Planner and reactive agent decide first whether the conversation answers the question fully, partly, or not at all, and retrieve only what is missing - Questions about the current or latest state still retrieve - Answer step reads the recent conversation alongside retrieved context - Planner and reactive agent see the most recent turns, not the oldest - Trace Logs mark a plan that answers without retrieval Refs: GML-2325
- Keep the MCP SDK below version 2, whose client API the add-on support does not use yet Refs: GML-2326
- Planner and reactive agent compare each fact the question asks for with the conversation - In planned mode, a plan without retrieval gets a second check - A question the conversation does not fully answer is looked up instead Refs: GML-2325
… rows - Structured query steps show the query that ran, plus failed or empty attempts - A structured query that returns no rows no longer counts as data - A valid openCypher query with no rows is not regenerated Refs: GML-2327
- Trace log of a failed answer ends with an Error step that states the error - Chat error message asks to rephrase the question, not the prompt Refs: GML-2328
- Structured queries go straight to openCypher when no functions are registered - A question that names no relationships no longer fails the function step Refs: GML-2327
- Drop the agent tools that ran self-written GSQL and neighbor expansion - GraphRAG code reaches TigerGraph only through pyTigerGraph Refs: GML-2332
- Planner gets a compact schema summary to decide on structural retrieval - The summary is cached per graph and refreshed when the schema changes Refs: GML-2331
- Classic and agentic engines judge a query result empty the same way - An empty result returned as JSON text counts as no rows in the classic engine Refs: GML-2327
- Connection tests no longer leave mocked modules behind for later tests - Tests follow per-user sign-in, three-part role lookup and masked secrets - Extractor tests use the schema-based extractor setup - LLM regression tests are skipped cleanly outside a git checkout Refs: GML-2333
- README describes schema-based planning and what Trace Logs record - CHANGELOG entries describe behavior without diagnostic detail
- Bump the version to 2.0.4 - Add the 2.0.4 entry to the README release list
- A plan summary or step reason that repeats a few characters in place is cut where the repeat starts Refs: GML-2324
- Gemini 3 planning, conversation check and triage use temperature 1.0 when the configured value is lower - Answers, knowledge graph building and other models keep the configured temperature Refs: GML-2335
- An earlier answer that says the information was not found, hedges, or is incomplete is retrieved again - Trace Logs list dropped plan steps and show the fallback retrieval - The fallback retrieval queries the graph as well as the documents Refs: GML-2325
…orrectly - A code block labelled opencypher or cypher no longer leaves the label in the query - Generated queries run in interpreted mode through pyTigerGraph, not the GSQL shell - Trace Logs show only the generated query, without the graph wrapper Refs: GML-2336
- A query that fails TigerGraph's check is not run; its errors go back to the generator - Up to 10 attempts (configurable), stopping early when the generator repeats itself - The generator uses TigerGraph's date functions, and queries that change data are rejected Refs: GML-2336
- Long unbroken lines in step input and output wrap inside the window - Download and Close stay visible at the bottom of the window Refs: GML-2327
- Answers without supporting information ask the user to rephrase with more specific information or scope - Fixed no-answer messages share one wording and no longer name internal steps Refs: GML-2328
- The planner reviews the conversation before planning and records the review in the plan - A plan without retrieval is used only when the review backs it; otherwise one replan, then the fallback retrieval - Plans must include their steps, so a plan can no longer come back empty - The separate conversation check is removed Refs: GML-2325
- Each request is recorded by its audit entry only; entry and HTTP client lines are at DEBUG - The web server's per-request access log is off Refs: GML-2337
- CHANGELOG entries state what changed since 2.0.3, not where problems were found Refs: GML-2323
- Structured query summaries give the row and character count, like document searches Refs: GML-2327
Collaborator
Author
|
@prinskumar-tigergraph please review the planner changes specifically for gemini-3.x |
- Schema given to query generation marks NULLABLE attributes - Generated queries use IS NULL only on NULLABLE attributes and compare other attributes with the default of their type - A query that fails to run because of a null test gets a rewrite hint - String comparisons lowercase both the attribute and the literal Refs: GML-2341
- The query a structured step ran is no longer shortened with its rows when the step output exceeds the trace size limit Refs: GML-2327
- The planner reviews its whole context, the conversation and the graph schema, before deciding to retrieve - Questions about the knowledge base itself are answered from the schema - The planner names the context and earlier turns the answer uses - An earlier answer is not reused for a question about the current state - The planner's schema names the graph the chat uses Refs: GML-2325
- Generated queries avoid combining a string match with OPTIONAL MATCH, CASE or a null test - A run error from such a combination names the construct to remove - Query generation stops when TigerGraph returns the same error three times in a row Refs: GML-2341
prinskumar-tigergraph
left a comment
prinskumar-tigergraph
left a comment
Contributor
There was a problem hiding this comment.
looks good for merge
- The planner's schema refreshes when a type description changes, not only when the schema changes - Query generation uses the current type descriptions on every query Refs: GML-2331
- Hold the OpenAI client below version 3 - Count a regenerated vertex only when its embedding is saved Refs: GML-2345
- Changelog entries state each problem and change, with ticket keys - Code comments name the constraint and the ticket instead of the cause Refs: GML-2323
Keys from the private tracker gave outside readers nothing to follow. - Changelog entries and code comments stand on their own - The openai version limit is marked as temporary, with its follow-up Refs: GML-2323
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Release GraphRAG 2.0.4 into
main. Merges all 2.0.4 work onrelease_2.0.4(based on currentmain).Epic: GML-2323
Added
cypher_max_attemptssetting — how many times a graph query may be generated for one question, default 10 (GML-2336).planner_max_output_tokenssetting — caps the planner's output per plan (GML-2324).Changed
Fixed
to_datetime()is used, each query is checked by TigerGraph before it runs, queries that change data are rejected (GML-2336), questions about missing values match the data, and generation stops when TigerGraph repeats the same error (GML-2341).Removed
tg_run_queryandtg_get_neighborsagent tools — custom planner or agent prompts that name them should drop them (GML-2332).Dependencies
openaiis held at>=2.45,<3(was>=1.92.2); 3.14.1 and later break Regenerate embeddings (GML-2345). The MCP SDK stays below 2 (GML-2326).Verification
Unit suite: 535 passed, 41 skipped, 0 failed (stale tests fixed in GML-2333). Live against a TigerGraph instance with gpt-5.4-mini and gemini-3.5-flash: multi-turn conversation tests in planned and reactive mode, and structured-query checks in planned, reactive and classic mode.
See
CHANGELOG.md[2.0.4] for full details.